Upload dataset

.gitattributes

@@ -57,3 +57,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+IPCC_AR6_WGIII_FullReport.txt filter=lfs diff=lfs merge=lfs -text

IPCC-61_decisions-adopted-by-the-Panel.txt

@@ -0,0 +1,1345 @@
+ 
+1  
+ 
+SIXT Y-FIRST SESSION OF THE IPCC  
+27 July – 2 August 2024 , Sofia, Bulgaria 
+ 
+Decisions adopted by the Panel  
+ 
+  
+Decision IPCC -LXI- 1.  Adoption of the Provisional A genda  
+Documents:  IPCC-L XI/Doc.1 and IPCC -LXI/Doc.1 , Add.1  
+ 
+The Intergovernmental Panel on Climate Change at its Sixty -first Session adopts the Provisional 
+Agenda as contained in document IPCC -LXI/Doc.1 .  
+ 
+ 
+ 
+Decision IPCC -LXI- 2.  Admission of Observer Organizations  
+ Document: IPCC-L XI/Doc. 3, Rev.1  
+ 
+The Intergovernmental Panel on Climate Change at its Sixt y-first Session decides to grant the following 
+organizations IPCC observer status, in accordance with the IPCC Policy and Process for Admitting 
+Observer Organizations:  
+1) Bureau international des poids et mesures (BIPM) 
+2) Children and Youth International (CYI)   
+3) Save the Climat e 
+4) Central American Commission on Environment and Development (CCAD) 
+5) International Society of City and Regional Planners (ISOCARP) 
+6) International Organization for Standardization (ISO)  
+7) Woodwell Climate Research Center (Woodwell)  
+8) Wellcome Trust (Wellcome)  
+9) West African Science Service Centre on Climate Change and Adapted Land Use (WASCAL)  
+10) Human Rights and Environment Improvement Centre (HREIC)  
+11) The Degrees Initiative (Degrees)  
+12) Coalition Climat pour la Biodiversité et le Développement (CCBD ) 
+ 
+ 
+Decision IPCC -LXI- 3. Ad Hoc Group on  Lessons Learned from the sixth assessment cycle  
+ Document: IPCC-L XI/Doc. 9  
+ 
+The Intergovernmental Panel on Climate Change at its Sixty -first Session appreciates and takes note 
+of the work of the Ad Hoc Group on Lessons Learned from the Sixth Assessment Cycle but also notes 
+that this work  does not reflect Panel consensus and the topics are indicative, not exhaustive. These 
+topics may be further discussed during the seventh assessment cycle in an inclusive and transparent 
+manner within the  IPCC as appropriate.   
+ 
+ 
+ 
+
+ 
+2 Decision IPCC -LXI- 4. Matters related to other IPCC activities – IPCC Scholarship Programme    
+ 
+Document: IPCC-L XI/Doc. 8 
+ 
+The Intergovernmental Panel on Climate Change at its Sixty -first Session agrees to the amendment of 
+the IPCC Scholarship Programme Trust Deed as to the election of a Chair of the Board of Trustees, 
+and accordingly requests the IPCC Secretariat to present t he amendment to the Trust Deed for the 
+Panel’s approval at the Sixty -second Session of the IPCC.  
+ 
+ 
+ 
+Decision IPCC -LXI- 5. Seventh assessment report (AR7) products – Outline of the Special 
+Report on Climate Change and Cities  
+ 
+Document: IPCC- LXI/Doc. 2, Rev. 1  
+ 
+The Intergovernmental Panel on Climate Change at its Sixty -first Session decides:   
+ 
+(1) To agree on the outline of the Special Report on Climate Change and Cities as contained in Annex 
+1 to this document.  
+ 
+(2) That the time schedule for the production of the Special Report is as follows:  
+ 
+9 August – 20 September 2024   Call for nominations of authors  
+23 September – 19 December   Selection of authors  
+10–15 March 2025  First Lead Author Meeting  
+21–25 July 2025   Second Lead Author Meeting  
+17 October – 12 December  2025  Expert Review of the First Order Draft  
+12–16 January  2026  Third Lead Author Meeting  
+8 May – 3 July  2026  Government and Expert Review of the Second Order 
+Draft  
+3–7 August  2026  Fourth Lead Author Meeting  
+11 December 2026 – 5 February 2027  Final Government Distribution of the Final Draft and  
+Government Review of the Summary for 
+Policymakers  
+15–19 March  2027  Approval of the Summary for Policymakers and 
+acceptance of the Special Report  
+ 
+ 
+(3) That the budget for the production of the Special Report is as contained in Decision IPCC -LX-10 
+on the IPCC Trust Fund Programme and Budget.  
+ 
+ 
+  
+  
+
+ 
+3  
+ANNEX 1  
+ 
+IPCC Special Report on Climate Change and Cities  
+ 
+ 
+Summary for Policymakers 
+Technical Summary  
+ 
+ 
+Chapter 1: Cities in the context of climate change: framing of the report  
+ 
+• Integrated storyline of the report, chapter narrative, sequence, and linkages to other relevant 
+processes and assessments  
+  
+• Framing and defining urban systems and settlements, and their regional and climatic characteristics (including complex, cascading, compounding, and repeating risks)  
+  
+• Sustainable development and climate resilience, acknowledging the diversity of development 
+status of cities and countries  
+  
+• Cities as hotspots of effects of hazards and emissions, losses and damages, vulnerabilities, exposure, and impacts, while also being key climate actors  
+  
+• Framing of multi -dimensional urban characteristics, including physical, socioeconomic and 
+environmental features  
+  
+• Treatment of urban vulnerabilities, marginalized areas and people, gender, equity, informality and justice  
+  
+• Psychology, perception, behaviour and attitudes toward climate change and cities  
+  
+• Interconnection between local context and global context (governance, science, and climate 
+change), and between urban and rural systems  
+  
+• Assessment methodologies, including following a regional approach, diverse knowledge systems (including Indigenous Knowledge), practitioner expertise, city networks, and considered time 
+frames and spatial scales  
+  
+Chapter 2: Cities in a changing climate: trends, challenges and opportunities  
+  
+• Understanding and learning from the past (global climate, hazards, crises, socioeconomic 
+developments); past, current and future global and city -specific climate (trends, means, extremes)  
+ 
+• Urbanization, urban service, common and different urban development trends (population, 
+demographics, informality and inequity, development stage, land use, geography, minorities and 
+intersectionality, urban extent, form, path dependencies, lock -in, retreat, reconstruction, growth 
+and decline, resource and carbon footprint, health and wellbeing, waste management, ecosystems, economy, finance and insurance, work, artificial intelligence and digitalization)  
+ 
+• Urban emissions trends including consumption- based emissions; the role of cities in emissions 
+and mitigation; future global and city -level scenarios, considering local options, equity, sustainable 
+development, infrastructure, and informal settlements  
+ 
+
+ 
+4 • City-specific risks and their global and regional climatic impact -drivers (extremes and their 
+attribution, slow -onset events, e.g., sea level rise); compounding and cascading risks; scenarios 
+with and without risk reduction, adaptation, resilience building, changes in vulnerability and 
+exposure across systems and sectors, including eco- systems and biodiversity, food, health and 
+housing, innovative technologies/methods (measurements and models)  
+  
+• Current mitigation and adaptation, planned and unplanned relocation, losses and damages 
+experienced, and the socio- economic trends that shape them, including policy, governance, 
+colonization   
+• Understanding the two- way interaction/feedback between cities, regions and countries, science 
+behind the interactions (understanding the biophysical mechanisms); social interactions; climate 
+and air quality, and other environmental changes, multi -hazard components (compounding and 
+cascading hazards)  
+  
+• Data, information, tools accessibility/availability/usability/transparency  
+  
+• Uncertainties, implementation gaps, unprecedented situations  
+  
+• Complexity and the need to contextualized climate change within broader societal trends 
+(geopolitical, polarizing societal trends) and goals (Sustainable Development Goals), justice, 
+cascading effects on critical infrastructure  
+  
+  
+Chapter 3: Actions and solutions to reduce urban risks and emissions 
+ 
+• Common and context specific urban mitigation options for spatial planning, energy (heating, 
+cooling, electricity), existing and new buildings and infrastructure, mobility and transport, water, 
+land, food, demand -side measures and behavioral change and cros s-sectoral, integrated 
+approaches in urban systems such as circularity  
+  
+• Common and context specific urban adaptation and disaster risk reduction options for managing risks in natural, ecological and human systems (including but not limited to physical infrastructure, 
+urban nature -based solutions and ecosystem -based adaptation,  and planning and social policies 
+such as relocation, health systems, early warning systems)  
+  
+• Evaluation of city actions across mitigation and adaptation, and responding to losses and 
+damages such as reconstruction and rehabilitation, including lessons -learned, effectiveness and 
+feasibility, mitigation measures with baseline emissions inventories and targets adopted by cities  
+ 
+• Urban observation and modelling tools for monitoring and evaluation for sectors and unaccounted sources  
+  
+• Local risk assessments using scientific information, Indigenous Knowledge, and local knowledge 
+of impacts, types and scales of adaptation responses (including positive experiences and 
+outcomes, and aspects of maladaptive practices) and adaptation cycles in various regions and 
+contexts  
+  
+• Integrating mitigation and adaptation into sustainable development and just transitions, planning 
+approaches under and for uncertainty, synergies and trade- offs, nexus approaches, social 
+innovation, climate resilient development, adaptation targets and the  role of cities in net -zero 
+targets  
+ 
+• Metrics for assessing mitigation and adaptation options in the context of sustainable development and the characteristics of and within cities, including service provisioning that delivers health and 
+well-being for all  
+  
+
+ 
+5 • Case studies/best practices/stories related to climate resilient development, adaptation, 
+decarbonization and low -carbon development in a diverse range of cities  
+  
+  
+Chapter 4: How to facilitate and accelerate change 
+ 
+• New ways of planning under and for uncertainty; the likelihood of tipping points  
+ 
+• Providing climate and information services to enable action, including evaluation of mitigation, 
+adaptation, responses to losses and damages, and the cost and benefits of action and inaction, 
+and sustainable development  
+ 
+• Innovation in governance, urban planning policies, decision- making, technology, urban service 
+provision, energy access and shelter, infrastructure, social systems, and finance, including 
+adoption of innovation, facilitation of societal trends, acknowledging the diverse capacities  
+ 
+• Institutional capacities, competencies, inclusive multi -level governance  
+  
+• Indigenous Knowledge, local knowledge, diverse knowledge systems and values  
+  
+• Policies for behavioural  and lifestyle changes including demand- side mitigation measures, 
+education for empowerment, community engagement, social movements and communications  
+  
+• Finance, financial instruments, legal frameworks, economic and policy instruments  
+  
+• Holistic planning and systems thinking approach towards decarbonized and climate resilient cities  
+  
+• Structural inequity, gender, colonialism, and justice 
+  
+• Enabling conditions for poverty eradication, equity in just transitions  
+  
+• Political will and leadership   
+• Conflicting goals and trade- offs 
+  
+  
+Chapter 5: Solutions by city types and regions 
+ 
+This chapter contains a synthesis of solution- relevant information and a collection of case studies by 
+city types in the context of urban sustainable development, distinguished by multi -dimensional 
+characteristics such as:   
+ 
+• Geographical location (regions)  
+ 
+• Development stage  
+ 
+• Informality  
+  
+• City climate and projections  
+  
+• Climatic impact -drivers  
+  
+• Adaptation and mitigation options  
+  
+• Sectoral contributions to the economy  
+ 
+
+ 
+6 • Migration, urbanization and demographic trends  
+ 
+• Fragility and conflict situations  
+  
+• Losses and damages, vulnerability, impacts and risks  
+ 
+• Early warning systems  
+  
+• Capacities  
+  
+• Inclusiveness, equity and justice  
+  
+• Governance  
+  
+• Climate finance  
+  
+  
+Annex I: Glossary  
+ 
+  
+ 
+  
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+ 
+   
+ 
+   
+ 
+  
+
+ 
+7  
+ 
+Decision IPCC -LXI- 6. Options for Expert Meetings and Workshop for the seventh assessment 
+cycle  
+ 
+Document s: IPCC-L XI/Doc. 7; IPCC -LXI/Doc. 7, Add. 1  
+ 
+The Intergovernmental Panel on Climate Change at its Sixt y-first Session: 
+ 
+Invite s the Bureaus of the Working Groups/TFI and the IPCC Chair to bring forward proposals for Expert 
+Meetings and Workshops at the Sixty -second Session of the IPCC (IPCC- 62) and  future IPCC sessions, 
+in line with Appendix A, paragraph 7.1 of the IPCC Principles and Procedures, taking into account the 
+views expressed by Member governments at the Sixty -first Session (I PCC-61) regarding document 
+IPCC- LXI/Doc. 7.  
+    
+
+ 
+8  
+ 
+Decision IPCC- LXI-7. Seventh assessment report (AR7) products – Outline of the 2027 IPCC 
+Methodology Report on Inventories for Short -Lived Climate Forcers  
+ 
+Document: IPCC-L XI/Doc. 6 
+ 
+The Intergovernmental Panel on Climate Change at its Sixty -first Session  decides:  
+ 
+(1) To prepare a Methodology Report with the following title”  2027 IPCC Methodology Report on 
+Inventories for Short -lived Climate Forcers ”; 
+ 
+(2) To agree on the Terms of Reference for the production of  a Methodology Report as contained 
+in Annex 1, the T able of C ontents as contained in Annex 2, the Instructions to Experts and 
+Authors as contained in Annex 3,  the Workplan as contained in Annex 4, each annex as attached  
+to this Decision ; and  
+ 
+(3) That the budget for the production of the Methodology  Report is as contained in Decision 
+IPCC- LX-10 on the IPCC Trust Fund Programme and Budget.  
+ 
+   
+ 
+    
+ 
+   
+ 
+    
+ 
+   
+ 
+    
+ 
+   
+
+ 
+9  
+Annex 1. Terms of Reference 
+ 
+2027 IPCC Methodology Report on Inventories for Short -lived Climate Forcers  
+Background  
+1. At the 49th Session (IPCC -49) held in May 2019 (in Kyoto, Japan) the IPCC approved the Task Force on 
+National Greenhouse Gas Inventories (TFI) to produce an IPCC Methodology Report on SLCFs following 
+the Appendix A to the Principles Governing IPCC Work (Decision IPCC -XLIX- 7). 
+2. IPCC TFI carried out preparatory work including Expert Meetings1 during the AR6 cycle. The Scoping 
+Meeting produced the draft Table of Contents, which is outlined in Annex 2.  
+Scope  
+3. The new Methodology Report will provide guidance on SLCF emissions which are:  
+- Anthropogenic, not including secondary human- induced substances  
+- National  
+- Annual  
+- Reported in mass units for each individual emitted species.  
+4. Coverage:  
+- Taking into account that this work aims to cover all IPCC inventory sectors with categories where the 
+science is assessed to be robust enough to provide guidance for a Tier 1 methodological approach 
+and have a relative contribution to the global/regional emissions of the species, species2 assessed 
+and potentially covered by the new Methodology Report will be NO X, CO, NMVOCs, SO 2, NH 3, BC 
+and OC, as well as emissions of primary particulate matter relevant for radiative forcing, as appropriate.  
+- Methane and halogenated species under Montreal Protocol and Kigali Amendment will not be 
+covered since these are already addressed by the 2006 IPCC Guidelines for National Greenhouse 
+Gas Inventories ( 2006 IPCC Guidelines ), the 2013 Supplement to the 2006 IPCC Guidelines for 
+National Greenhouse Gas Inventories: Wetlands ( Wetlands Supplement ) and the 2019 Refinement 
+to the 2006 IPCC Guidelines for National Greenhouse Gas Inventories ( 2019 Refinement ). 
+- For NMVOCs, the methodology should provide estimates for total NMVOCs. The speciation of NMVOCs should be considered by authors, as appropriate.  
+- Anthropogenic emissions
+3 only, where anthropogenic refers to emissions from human activities and 
+from managed4 land.  
+- Sources covered are those of anthropogenic emissions, where scientific evidence is available; while for others, guidance could be provided as a basis for future methodological development.  
+- Geographical and temporal coverage is national and annual level, and authors should also consider guidance on spatial and temporal disaggregation of SLCF emissions.  
+5. Key elements:  
+- Structure: Information on each sector will be synthesised into a single document (a volume for each 
+of the inventory sectors: Energy, Industrial Process and Product Use (IPPU), Agriculture, Forestry 
+and Other Land Use (AFOLU), Waste. There will also be a v olume on cross -cutting issues, including 
+reporting tables).  
+ 
+ 
+ 
+1  The Joint 1st and 2nd IPCC Expert Meeting on SLCFs: https://www.ipcc -nggip.iges.or.jp/public/mtdocs/2110_SLCF.html  
+   The 3rd IPCC Expert Meeting on SLCFs: https://www.ipcc -nggip.iges.or.jp/public/mtdocs/2204_SLCF_EM3.html  
+2  Given  the uncertainties in the radiative forcing of H 2 and taking note that H 2 has not yet been well assessed as a climate forcer by IPCC 
+WGI, H 2 emissions relevant for radiative forcing are to be considered by the authors as  an Appendix subtitled “Basis for future 
+methodological development”  subject to the IPCC’s Principles and Procedures on review and adoption.  
+3  as defined in the 2006 IPCC Guidelines for National Greenhouse Gas Inventories ( 2006 IPCC Guidelines ), the 2013 Supplement to the 
+2006 IPCC Guidelines for National Greenhouse Gas Inventories: Wetlands ( Wetlands Supplement ) and the 2019 Refinement to the 
+2006 IPCC Guidelines for National Greenhouse Gas Inventories ( 2019 Refinement ). 
+4   land where human interventions and practices have been applied to perform production, ecological or social functions.  
+
+ 
+10 - Content of cross -cutting guidance: The volume for cross -cutting issues will include: introduction5, with 
+guidance on SLCF species and definitions, approaches to data collection6; uncertainties; 
+methodological choice and identification of key categories; time series consistency; quality 
+assurance/quality control (QA/QC) and verification; and reporting guidance and tables.  
+- Content of sectoral guidance: The volumes for each sector will include tiered methodological 
+approaches; decision trees; methods and emission factors, where appropriate; cross -references as 
+necessary to avoid double counting or omissions of emissions; sect or-specific guidance on 
+uncertainty assessment and QA/QC; and reporting and documentation guidance.  
+Approach  
+6. The result of the work will be an IPCC Methodology Report “2027 IPCC Methodology Report on Inventories for Short -lived Climate Forcers”.  
+7. The authors will ensure consistency with categories and build on the   methodological guidance within the 
+2006 IPCC Guidelines, Wetlands Supplement  and 2019 Refinement .  
+8. The authors  will follow “Instructions to Experts and Authors” presented in Annex 3 to ensure a consistent 
+and coherent approach across all the volumes and chapters, including the use of common terminology.  
+9. Importantly, the authors will provide guidance based on the good practice
+7 guidance definition and the 
+structured tiered approach described in the 2006 IPCC Guidelines, Wetlands Supplement and 2019 
+Refinement.   
+10. The production of the Methodology Report will be completed in 2027 as noted in the work plan in Annex 4 following Decision IPCC -LX-9. 
+  
+ 
+5  considering the importance for climate effects of spatial distribution and temporal resolution of SLCF emissions, and changes  in co - 
+emitted species  
+6  including generic methods of measurements, approaches to estimate BC/OC , including on techniques of measurement and all variables 
+used to derive emission factors,  NMVOC speciation, spatial distribution and temporal resolution, technology, and abatement information.  
+7 "Good practice" is a key concept for inventory compilers to follow in preparing national greenhouse gas inventories. The key concept does not change 
+in the 2019 Refinement. The term "good practice" has been defined, since 2000 when this concept was introduced, as "a set of procedures intended to 
+ensure that greenhouse gas inventories are accurate in the sense that they are systematically neither over - nor underestimates so far as can be judged, 
+and that uncertainties are reduced so far as practicable". This  definition has gained general acceptance amongst countries as the basis for inventory 
+development and its centrality has been retained for the 2019 Refinement. Certain terms in the definition have been updated b ased on feedback from 
+the statistics communi ty, such that this definition can be also understood as "a set of procedures intended to ensure that greenhouse gas inventori es are 
+accurate in the sense that they are systematically neither over - nor underestimates so far as can be judged, and that they are precise so far as 
+practicable" in the context of refinement of Chapter 3 of Volume 1.  
+Good Practice covers choice of estimation methods appropriate to national circumstances, quality assurance and quality control at the national level, quantification of uncertainties and data archiving and reporting to promote transparency.  
+ 
+ 
+
+ 
+11 Annex 2. Table of Contents  
+ 
+2027 IPCC Methodology Report on Inventories for Short -lived Climate Forcers  
+ 
+Overview  
+Volume 1. General Guidance  
+ Introduction  
+(including, but not limited to: Background on SLCFs and their importance for climate, Key differences 
+between SLCFs and GHGs emissions, Holistic approaches to SLCFs and the importance of co- emitted 
+species, Spatial distribution and temporal resolution and relevance to climate effects, Interlinkages with meteorology, Importance of technologies and abatement technologies)  
+ Approaches to Data Collection  
+(including, but not limited to: Spatial distribution and temporal resolution, Measurement techniques, NMVOC speciation, Technologies and Abatement technologies)  
+ Uncertainties  
+ Methodological Choice and Identification of Key Categories  
+(including, but not limited to KCA by SLCF species, Issues of co- emitted species in SLCF KCA)  
+ Timeseries consistency  
+(including, but not limited to: Addressing changes in measurement techniques, Addressing changes in technologies, including for abatement)  
+ QA/QC and Verification  
+(including, but not limited to: Consistency with co- emitted GHGs and SLCFs, Comparison with 
+global/regional inventories, Comparisons with atmospheric observations and models)  
+ Reporting guidance and Tables  
+Volume 2. Energy Sector  
+ Introduction  
+ Stationary combustion  
+ Mobile combustion  
+ Fugitive Emissions  
+ Other  
+Volume 3. IPPU Sector  
+ Introduction  
+ Mineral Industry  
+ Chemical Industry  
+ Metal Industry  
+ Non- Energy products from fuels and Solvent  Use 
+ Other  
+Volume 4. AFOLU Sector  
+ Introduction  
+ Generic methodologies  
+ Consistent representation of land  
+ Emissions from Livestock and Manure Management  
+ Land use categories  
+ Managed soil
+8 
+ Other  
+Volume 5. Waste Sector  
+ Introduction  
+ Solid Waste Disposal  
+ Biological Treatment of Solid Waste  
+ Incineration and Open Burning of Waste  
+ Wastewater Treatment and Discharge  
+ Other  
+ 
+  
+ 
+8 As expanded by the Wetlands Supplement guidance/categorization  
+
+ 
+12 Annex 3.  Instructions to Experts and Authors  
+ 
+2027 IPCC Methodology Report on Inventories for Short -lived Climate Forcers  
+ 
+1. Work on a 2027 IPCC Methodology Report on Inventories for Short -lived Climate Forcers will be guided by 
+the IPCC procedures for the Preparation, Review, Acceptance, Adoption, Approval and Publication of the 
+IPCC Reports (Appendix A to the Principles Governing the IPCC Work9). This document is consistent with 
+the IPCC procedures and applies to all experts engaged in the production of a new Methodology Report.  
+2. In this document the term “experts” covers Co- Chairs, members of the TFI Bureau (TFB), technical support 
+unit (TSU) Staff, Coordinating Lead Authors (CLAs), Lead Authors (LAs), and Review Editors (REs) as well as Contributing Authors (CAs) and Expert Revie wers.  
+3. These notes are intended as guidance to experts contributing to a new Methodology Report . They are 
+intended to ensure a consistent and coherent approach across all the volumes or chapters and to promote 
+common terms used.  
+Confidentiality  
+4. Authors meetings  are closed meetings. Any discussions are confidential except for any published report of 
+the meeting. This is to ensure that experts participating in the meetings can express themselves and discuss 
+issues freely and openly.  
+5. The IPCC considers the drafts of a new Methodology Report , prior to acceptance, to be pre- decisional, 
+provided in confidence to reviewers, and not for public distribution, quotation or citation.  
+6. The TSU will keep drafts of a new Methodology Report  sent for the IPCC review, any comments received 
+on them and the responses by authors. All written expert and government review comments will be made available to reviewers on request. These will be made available on the IPCC website as soon as possible after the acceptance by the Panel and the finalisation of the report.
+  
+Conflict of Interest  
+7. It is important that all experts involved in the IPCC activities avoid any conflict of interest or the direct and 
+substantial appearance of a conflict of interest. It is recognised that many experts in Emission Inventories 
+are employed by, or funded by, parties with some interest in the outcome (e.g. most inventory compilers 
+are funded by national governments or industry). It is therefore important to be open and transparent about financial and other interests.  
+8. The IPCC implements a Conflict of Interest (COI) Policy
+10  that applies to all individuals directly involved in 
+the preparation of IPCC reports, including senior IPCC leadership (IPCC Chair and Vice- Chairs), other 
+Bureau and Task Force Bureau members, authors with responsibilities for report content (CLAs, LAs),  
+Review Editors and staff of the TSU. The overall purpose of this policy is to protect the legitimacy, integrity, 
+trust, and credibility of the IPCC and of those directly involved in the preparation of reports, and its activities.  
+9. Before an individual is appointed as a CLA, LA and RE for a new Methodology Report , the TFB will request 
+the individual to complete a Conflict of Interest Disclosure Form (“the COI Form”) contained in Annex B to 
+the COI Policy which will be submitted to the TSU. The TFB will then evaluate the form to determine whether 
+the individual has  a conflict of interest that cannot be resolved.  
+10. All CLAs, LAs and REs will inform the TSU annually of any changes in the information provided in their 
+previously submitted COI Form. The TFB will evaluate the revised information.  
+11. All COI Forms and any records of the deliberations of the COI Expert Advisory Group, deliberations and/or 
+decisions of the COI Committee in relation to conflict of interest issues in respect of specific individuals and 
+any information disclosed by individuals for the purposes of the COI Policy will be transferred to the 
+Secretariat after they have been reviewed and will be securely archived by the Secretariat and retained for 
+a period of five years after the end of the assessment cycle during which the relevant individual contributed, 
+after which the information will be destroyed. Subject to requirement to notify the existence of a conflict of interest to others, the information referred to above will be considered confidential and will not be used for 
+any p urpose other than consideration of conflict of interest issues under these Implementation Procedures 
+without the express consent of the individual providing the information.  
+ 
+ 
+ 
+9 https://www.ipcc.ch/site/assets/uploads/2018/09/ipcc -principles -appendix -a-final.pdf  
+10 https://www.ipcc.ch/site/assets/uploads/2018/09/ipcc -conflict -of-interest -2016.pdf  
+
+ 
+13 Responsibilities of authors and other experts  
+12. The role of authors is to impartially assess ALL the available literature and to describe the best 
+methodologies available. Experts should be impartial. Authors should review all literature available up to a 
+cut-off date to be decided by the TFB as part of  the agreed work plan.  
+13. After drafting the report authors will be asked to consider all comments received on the drafts and to adjust 
+and revise the text accordingly. They should document their responses. If they do not accept a comment this should be explained. Review Editors should check whether the accepted changes were fully 
+incorporated in the revised text.  
+14. Responsibilities and duties of authors and other experts are currently explained in more detail in the IPCC 
+procedures for the Preparation, Review, Acceptance, Adoption, Approval and Publication of the IPCC Reports (Appendix A to the Principles Governing the IPCC Work).  
+Literature  
+15. The use of literature should be open and transparent. In the drafting process, emphasis is to be placed on 
+the assurance of the quality of all cited literature. Priority should be given to peer -reviewed scientific, 
+technical and socio- economic literature i f available.  
+16. It is recognized that other sources provide crucial information for IPCC Reports. These sources may include 
+reports from governments, industry, and research institutions, international and other organizations, or 
+conference proceedings. Use of this literat ure brings with it an extra responsibility for the author teams to 
+ensure the quality and validity of cited sources and information as well as providing an electronic copy. In general, newspapers and magazines are not valid sources of scientific information. Blogs, social networking 
+sites, and broadcast media are not acceptable sources of information for IPCC Reports. Personal 
+communications of scientific results are also not acceptable sources.  
+17. For any sources written in a language other than English, an executive summary or abstract in English is required.  
+18. All sources will be integrated into a reference section of an IPCC Report.  
+19. For more details of the procedure on the use and referencing of literature in IPCC Reports, see Annex 2 to 
+the IPCC procedures for the Preparation, Review, Acceptance, Adoption, Approval and Publication of the IPCC Reports (Appendix A to the Principles Governing the IPCC Work).  
+Principles of the new Methodology Report  
+20. Guidance in the new Methodology Report should be understandable and easy to implement. Lead authors 
+should make efforts to balance the need to produce a comprehensive self -contained report with reasonable 
+limits to the length and detail of the guidance. In particular:  
+a. The guidance should follow a cookbook approach by providing clear step by step instructions. It 
+should not try to be a textbook. Detailed background information on emission processes, scientific 
+studies, etc. is generally referenced rather than included.  
+b. Lead authors must consider relevant scientific developments and national methods used by countries in their inventories.  
+c. Authors should bear in mind that the target audience is a diverse group of readers who are primarily concerned with the elaboration of national inventories.  For this reason, the emphasis should be on ensuring clear communication of practical and understandable guidance.  
+21. This work aims to cover all IPCC inventory sectors with categories where the science is considered to be robust enough to provide guidance for a Tier 1 methodological approach and have a relative
+11 contribution 
+to the global/regional emissions of the species, using the significance and prioritization criteria as shown 
+below.  
+Significance and prioritization criteria  
+• Significance of the category and the species within the sector on a global/regional scale. Categories  
+significant only for a limited number of particular countries, currently or in the foreseeable future, 
+may not meet this criterion.  
+• Sufficient data availability and maturity of scientific advances to provide a basis for methodological 
+development, including:  
+o Ability to develop default emission factors and parameters  
+o Feasibility of obtaining the necessary data to implement the methods  
+ 
+ 
+11 i.e. not insignificant  
+
+ 
+14 22. The general structure, approach and definitions used in the 2006 IPCC Guidelines , such as tiered approach 
+and decision trees will be followed. Annexes may be used where necessary to contain additional data to 
+support the methodologies, although large numbers of annexes will probably not be necessary. Appendices are not ruled out where scientific knowledge is insufficient for countries to agree full methodologies, but 
+please avoid as far as possible work on areas that have to be relegated to an appendix. Appendices should be sub- titled by “Basis for future methodological development”.  
+Definitions  
+23. The following terms will be used throughout the new Methodology Report, and it is essential that all Lead Authors have a common understanding of their meaning and relevance.  
+24. Tier A Tier refers to a description of the overall complexity of a methodology and its data requirements. 
+Higher tier methods are generally more complex and data- intensive than lower tier methods. The guidance 
+for each category should contain at least a Tier 1 method, and in many cases there will be a Tier 2 and Tier 
+3. The general expectation is that Tier 2 and Tier 3 methods will both be consistent with good practice  
+guidance for key sources, although in some cases Tier 3 will be preferred.  
+25. T ier 1  approaches are simple methods that can be applied by all countries in all circumstances. Default 
+values for the emission factors and any other parameters needed must be supplied (see below for documentation needed).  
+26. T
+ier 2  methods should in principle follow the same methodological approach as Tier 1 but allow for higher 
+resolution country specific emissions factors and activity data. In some categories, this may not be the case. These methods should better replicate the parameters affecting the emissions. Country specific emission factors are needed and possibly more parameters will also be needed.  
+27. T
+ier 3  methods give flexibility either for country specific methods including modelling or direct measurement 
+approaches, or for a higher level of disaggregation, or both. This is a more complex method, often involving a model. This will replicate many features of nation emissions and require specific parameters for each country.  
+28. D
+efault information  is data that is appropriate for use where there is no better detailed, country specific 
+information. If appropriate, authors may specify regional default data. Users of the guidelines should be 
+encouraged to try to find better country specific data. Default data are appropriate for Tier 1 methods and 
+the guidelines should contain all the default values needed. Emission factors for higher tiers need not be specified because it is a function of higher tier methods to find data reflecting national circumstances. Default information is included primarily to provide users with a starting point from which they can develop their own national assumptions and data. Indeed, national assumptions and data are always preferred because the default assumptions and data may not always be appropriate for specific national contexts. In 
+general, therefore, default assumptions and data should be used only when national assumptions and data 
+are not available.  
+29. D
+ecision Trees. A decision tree is a graphical tool to assist countries in selecting from the IPCC methods.   
+30. Key categories  are inventory categories which individually, or as a group of categories (for which a 
+common method, emission factor and activity data are applied) are prioritised within the national inventory system because their estimates have a significant influence on a country’s total inventory in terms of the 
+absolute level, the trend, or the level of uncertainty in emissions. Key category analysis should be performed 
+species by species. The appropriate threshold to define key categories should be considered by auth ors. 
+31. S
+ector  refers to the four sectors of the guidelines (Energy; Industrial Process and Product Use (IPPU); 
+Agriculture, Forests and Other Land Use (AFOLU) and Waste) these are divided into categories and subcategories.  
+a. Sector 1  
+b. Category 1.A  
+c. Sub-category 1st order 1.A.1 
+d. Sub-category 2nd order 1.A.1.a  
+e. Sub-category 3rd order, 1.A.1.a.i  
+32. Worksheets . These will be printed versions of spreadsheet tables, that, when filled in, enable the user to 
+perform the emission estimation. They should contain all the calculations and written text with any formulae. Additional worksheets may be required to compile the results of the worksheets into the reporting tables.  
+33. R
+eporting Tables  are tables that present the calculated emission inventory and sufficient detail of other 
+data used to prepare the inventories for others to understand the emission estimates.  
+ 
+  
+
+ 
+15 34. Usage:  
+a. “Good Practice”  is defined in the 2019 Refinement as follows: “a key concept for inventory compilers 
+to follow in preparing national greenhouse gas inventories. The key concept does not change in the 
+2019 Refinement. The term "good practice" has been defined, since 2000 when this concept was 
+introduced, as "a set of procedures intended to ensure that greenhouse gas inventories are accurate 
+in the sense that they are systematically neither over - nor underestimates so far as can be judged, 
+and that uncertainties are reduced so far as practicable". This definition has gained general acceptance amongst countries as the basis for inventory development and its centrality has been retained for the 2019 Refinement . Certain terms in the definition have been updated based on 
+feedback from the statistics community, such that this definition can be also understood as "a set of 
+procedures intended to ensure that greenhouse gas inventories are accurate in the sense that they 
+are systematically neither over - nor underestimates so far as can be judged, and that they are precise 
+so far as practicable" in the context of refinement of Chapter 3 of Volume 1”.  
+The concept mentioned above should be applied to all species dealt with in this report.  
+b. Good Practice covers choice of estimation methods appropriate to national circumstances, quality 
+assurance and quality control at the national level, quantification of uncertainties and data archiving and reporting to promote transparency.  
+c. “Shall ” should not be used. Either say “Good Practice is…” or say what needs to be done or what 
+should be done. These all indicate what needs to be done to comply with Good Practice.  
+d. "B e encouraged to" indicates a step or activity that will lead to higher quality inventory but are not 
+required for ensuring consistency with the IPCC Guidelines.  
+e. “R ecommend ” should not be used. In the GPG2000, the word “recommend” was avoided and 
+“Suggested” was used instead.  
+f. “ I nventory agency” is the body responsible for actually compiling the inventory, perhaps from 
+contributions from a number of other bodies while “ inventory compiler ” is the person actually 
+compiling the inventory,  
+Reporting Tables and worksheets  
+35. Worksheets reflect the application of tier 1 methods only, due to the varied implementation of higher tier methods by countries. Lead authors should stress the importance of documentation and archiving of particular types of information of relevance to each category, although advice may be given of what needs to be reported for transparency at higher Tiers.  
+Emission factors and methods  
+36. Authors should provide default emission factors and parameters. In doing this work, they should draw on the widest possible range of available literature, scientific articles and country reports.  
+37. All data reported in the guidance as IPCC default values shall be justified by authors by providing TSU with all background data used, and the source of those data, as well as all information on the method applied to derive the default values from the background data, as needed to replicate the calculation, in a timely manner as drafts are being developed. Background data should be compiled in the attached form (Appendix 1) to facilitate the upload in the Emission Factor Database (EFDB).  Lead authors should be familiar with 
+the draft cross -cutting guidance on data collection in Volume 1 and the guidance on cross -cutting issues in 
+this note on terms, data types, data demands of methods and stratification requirements.  Default data 
+should also meet the EFDB evaluation criteria – robustness, documentation, and applicability
+12.   
+38. Authors should develop guidance to provide additional information on rationale, references and background information on parameters used for estimating of default values where such information is available (similar 
+to Annexes in Chapter 10, Volume 4, of th e 2019 Refinement ), with a view to enhancing the transparency 
+and applicability of default values presented in the new Methodology Report.  
+39. Single IPCC default emission factors might not be ideal for any one country, but they can be recommended 
+provided that regional factors are unavailable, and the defaults are representative of typical conditions as far as can be determined.  It may be necessary or appropriate to provide a range of default emission factors along with clear guidance about how countries should select from within the range. Lead authors may also 
+provide multiple default emission factors, disaggregated by region, technology (including abatement 
+technologies), or another relevant classification scheme.  
+40. It is important to provide more default emission factors that reflect the unique conditions of developing 
+countries. In general, default emission factors for Tier 1 should represent emissions without category -
+specific mitigation measures, as well as relevant abatement technologies for which data are available. 
+ 
+ 
+12 EFDB evaluation criteria: https://www.ipcc -nggip.iges.or.jp/EFDB/documents/EFDB_criteria.pdf  
+
+ 
+16 41. Users of the guidelines should be encouraged to develop and use country specific data. Emission factors 
+for higher tiers need not be specified in the 2027 IPCC Methodology Report on Inventories for  Short -lived 
+Climate Forcers . Default information is included primarily to provide users with a starting point from which 
+they can develop their own national assumptions and data. Indeed, national assumptions and data are 
+always preferred because the default assumptions and data may not always be appropriate for s pecific 
+national contexts.  
+42. The basic principle concerning national methods will continue to apply –  countries are encouraged to use 
+national data or methods so long as they are consistent with the IPCC Guidelines.  
+43. Authors shall prefer IPCC methods applied to estimate GHG emissions when those can be straightforwardly 
+applied to estimate SLCF emissions as well as when those can be applied mutatis mutandis. The use of 
+consistent methodologies allows inventory -compilers  to use the same datasets for both sets of estimates. 
+This is to enhance efficiency in the use of resources available to inventory -compilers and thus to promote 
+accuracy of estimates.  
+44. Where the method applied for SLCF differs from that applied to estimate GHG emissions from the same 
+source, or the source is not covered in the 2006 IPCC Guidelines , in addition to methodological guidance, 
+guidance on activity data sources available at international level, and where possible at national level, will 
+be provided.  
+45. Authors should note the issue of double- counting, for example in the Energy sector the IPCC default method 
+for combustion assumes an Oxidation Factor equal to 1 resulting in all carbon calculated as CO 2, while the 
+addition of SLCF methods will require to estimate also other carbon compounds (CH 4, CO, NMVOC and 
+BC/OC). Authors should provide guidance to inventory compilers on how to address the issue of double-
+counting.  
+46. For BC/OC emissions, authors should provide guidance, including on techniques of measurement and all variables used to derive emission factors.  
+47. In considering the methodologies for SLCF emissions in the AFOLU sector, authors should not include 
+natural background emissions from land as these are not considered to be anthropogenic.  
+Boxes  
+48. Consistent with the 2006 IPCC Guidelines , the new Methodology Report may contain Boxes, which 
+should not be used to provide methodological guidance, but for information purposes or providing examples.  
+Decision trees  
+49. Consistent with the format and structure of the 2006 IPCC Guidelines , the new Methodology Report may 
+contain a decision tree for some sub- categories to assist countries in selecting from the IPCC methods. 
+These decision trees link the choice of IPCC methods to national circumstances via specific questions about 
+data availability and status as a key category
+13.  
+50. To ensure consistency in decision tree logic and format across categories, lead authors should adhere to the following requirements:  
+a. The decision trees should be based on a series of questions with clear yes/no answers, and two subsequent branches along yes/no paths.   
+b. The decision trees should start with assessing data availability for the highest tier method, and then direct countries step- wise towards lower tier methods if activity data, emission factors or other 
+parameters are not available.   
+c. The decision tree should indicate the lowest tier method that is judged to be appropriate for estimating 
+emissions from a key category.  
+d. If data are not available for the method referred to in c, the ‘No’ response should direct the reader to the question “Is this a key category?”  If the answer to this is ‘Yes’, the decision tree should 
+recommend that the country collect the necessary data to implement a higher tier method.  If the 
+answer is ‘No’, then the decision tree can recommend a lower tier method. There is no need to deal with the case for a key source where a country does not have the resources to gather additional data needed to implement higher Tier methods. This is dealt with in Volume 1 of the 2006 IPCC Guidelines .  
+e. The branches of the decision trees should end in ‘out -boxes’ that correspond to specific tiers identified 
+in the guidance for that category and are labelled by Tier.  Lead authors may also recommend out -
+boxes for hybrid tiers.  
+ 
+ 
+13 The most appropriate choice of estimation method (or tier) may also depend on national circumstances, including the availabil ity of 
+resources and advice on this will be given in the cross -cutting volume.  
+
+ 
+17  
+f. Lead authors may develop separate decision trees for different sub- categories.  Alternatively, they 
+may include decision tree options for selecting different tiers for different sub- categories.  This second 
+option is appropriate if it is advantageous to recommend a higher tier method only for significant sub-
+categories rather than for the entire category. Decision trees that use the ‘significance’ criterion 
+must include the “25 -30% rule”14, as reassessed by authors.  
+51. Additional Formatting Guidelines (see example):  
+a. Decision trees should be drafted in separate files. The TSU will integrate these files into the main text 
+at a later date.  
+b. Decision trees should NOT ask the question: “Does this source occur in the country?” This is because decision trees will only be used for sources which occur.  
+c. There should be a “START” box.  
+d. “Diamonds” should be used for questions/decisions.  
+e. “Squares” should be used for all other information.  
+f. The out -boxes should be individually numbered.  
+g. The text font should be Times New Roman 10pt.  
+h. Text should be centred within the boxes.  
+ 
+14 As defined in the 2019 Refinement  (i.e., a significant sub- category is one that makes up more than 25- 30% of emissions from a category).  
+
+ 
+18 Example. Decision tree for estimating emissions from fuel combustion  
+      (Figure 1.2 Chapter 1 Volume 2 of the 2006 IPCC Guidelines)  
+ 
+ 
+
+
+ 
+19 Units  
+52. SI units shall be used throughout: in text, equations, worksheets and tables.  Emissions have to be 
+expressed in mass units and units have to be used consistently within each sector. When similar activity 
+data is used for different sectors same units need to be used (CLAs have to take care about such 
+harmonisation). Conversion factors have to be provided (for example to estimate N 2O from N). Where input 
+data available may not be in SI units conversions should be provided.  
+53. Standard abbreviations for units and chemical compounds are given in Appendix 2.  
+
+ 
+20 Appendix 1. EFs and parameters Documentation  
+ 
+This form should be used to document all EFs and parameters used in the new Methodology Report. This gives 
+the minimum information that should be considered by the authors.  
+ 
+Author1   
+IPCC Category    
+Name of EFs / parameters   
+Activity, e.g. Fuel2 in the Energy 
+Sector    
+ Species3:  CO NOx  … … … 
+Value:       
+Unit:      
+Uncertainty (as +/% or 2.5 and 97.5 percentiles )
+4      
+Applicability5 – fill in as necessary 
+if data not generally applicable. 
+Describe appropriate 
+Technologies, Practices, 
+Abatement Technologies, Region, 
+and/or Regional Conditions    
+Source of data (chose one)  Measurement - Scientific Literature  
+Other Measurement  
+National Inventory Report  
+Calculated  
+Based on fuel quality  
+Expert Judgement6 
+Method of derivation of the value 
+(e.g., arithmetic mean, weighted 
+mean, adjustment of a literature 
+data by expert judgment etc.   
+Reference7   
+URL  
+Abstract in English (if the abstract is in another language)   
+ 
+Note:  
+1. The author is the LA/CA/CLA who writes the relevant section and proposes the data.  
+2. Fuels as defined in the Energy volume of the 2027 IPCC Methodology Report on Inventories for Short -
+lived Climate Forcers  
+3. Add additional species as required  
+4. As defined by cross -cutting volume  
+5. Only to be completed where it is necessary to specify the applicability of the data  
+6. Attach the elicitation protocol  
+7. As reference to document, report, calculation or if expert judgement to those involved (Names or group 
+e.g. “Waste BOG on Solid Waste Disposal Sites”) with DOI, where possible  
+
+ 
+21 Appendix 2. Units and Abbreviations  
+ 
+Abbreviations of, and how to spell, chemical species  
+BC  Black Carbon  
+CCl 4  Carbon tetrachloride  
+CF4  Tetrafluoromethane  
+C2F6  Hexafluoroethane  
+CFCs   Chlorofluorocarbons  
+CH 4  Methane  
+CO  Carbon monoxide  
+CO 2  Carbon dioxide  
+EC  Elemental Carbon  
+H2  Hydrogen  
+HFCs   Hydrofluorocarbons  
+NH 3  Ammonia  
+NMVOCs   Non- methane volatile organic compounds  
+NO X  Nitrogen oxides  
+N2O  Nitrous oxide15  
+OC  Organic Carbon  
+PFCs   Perfluorocarbons  
+PM x   Particulate Matter (x –  micrometres)  
+S  Sulphur  
+SF6  Sulphur hexafluoride  
+SO 2  Sulphur Dioxide  
+  
+ 
+15 In the IUPAC N 2O is officially named “Dinitrogen Oxide”. However, “nitrous oxide” is widely used and understood in the emission inventory 
+community and by the UNFCCC and so, to avoid confusion, will be used.   
+
+ 
+22 Units and abbreviations  
+cubic metre   m3 
+hectare   ha 
+gram   g 
+gigagram   Gg 
+tonne   t 
+gigatonne   Gt 
+joule   J 
+degree Celsius   ℃ 
+calorie   cal 
+year  Yr 
+capita   Cap 
+gallon   gal 
+dry matter   Dm 
+atmosphere   atm 
+ 
+Prefixes and multiplication factors  
+Multiplication Factor  Abbreviation  Prefix  Symbol  
+1 000 000 000 000 000  1015 peta P 
+1 000 000 000 000  1012 tera T 
+1 000 000 000  109 giga G 
+1 000 000  106 mega  M 
+1 000  103 kilo k 
+100 102 hecto  h 
+10 101 deca  da 
+0.1 10-1 deci d 
+0.01 10-2 centi  c 
+0.001  10-3 milli m 
+0.000 001  10-6 micro  μ 
+ 
+  
+
+ 
+23 Standard equivalents  
+1 tonne of oil equivalent 
+(toe)  1 x 1010 calories  
+103 toe  41.868 TJ  
+1 short ton   0.9072 tonne  
+1 tonne   1.1023 short tons  
+1 tonne   1 megagram  
+1 kilotonne   1 gigagram  
+1 megatonne   1 teragram  
+1 gigatonne   1 petagram  
+1 kilogram   2.2046 lbs  
+1 hectare   104 m2 
+1 calorie IT  4.1868 joule  
+1 atmosphere   101.325 kPa  
+  
+
+ 
+24 Annex 4. Workplan 
+ 
+2027 IPCC Methodology Report on Inventories for Short -lived Climate Forcers  
+ 
+Date  Action  Comments  
+February 2024  Scoping Meeting  Prepare ToR, ToC, Workplan and Guidance to authors  
+February 2024  TFB36 Meeting  Adoption of Outcomes of the Scoping Meeting and 
+Submission to IPCC  
+3rd quarter 2024  IPCC -61 IPCC Plenary approves ToR, ToC, Workplan and Guidance to authors  
+3rd quarter 2024  Call for Nomination of Authors and Review Editors  IPCC invites nominations from governments and international organizations  
+3rd quarter 2024  Establishment of the Steering Committee  TFB select members to join TFI Co- Chairs in the 
+Steering Group (to ensure consistency across all the volumes and continuity with the earlier IPCC inventory reports)  
+4th quarter 2024  Selection of Coordinating Lead Authors, Lead Authors 
+and Review Editors  Selection by TFB considering expertise and 
+geographical and gender balance  
+1st half 2025  1st Lead Author Meetings  LAM1 to develop zero order draft (ZOD)  
+2nd half 2025  2nd Lead Author Meeting  To develop first order draft (FOD) for review  
+1st quarter 2026  
+(8 weeks)  Expert Review  8 weeks review by experts  
+1st half 2026  Science Meeting  A small meeting of CLAs and some LAs to discuss specific issues that require intensive discussion to 
+reinforce the writing process  
+1st half 2026  3rd Lead Author Meeting  To consider comments and produce second order draft 
+(SOD) for review  
+2nd half 2026  Literature cut -off date (one 
+week before SOD Review)  Peer -reviewed papers accepted by the cut -off date 
+(even if not yet published) will be considered. Non- peer-
+reviewed documents which are made publicly available by the cut -off date.  
+2nd half 2026  
+(8 weeks)  Government & Expert Review  8 weeks review by governments and experts  
+1st half 2027  4th Lead Author Meeting  To consider comments and produce final draft (FD)  
+1st half 2027  Government Review  Distribute to governments for their consideration prior to approval (at least 4 weeks prior to the Panel)  
+2nd half 2027  Adoption/acceptance by IPCC  Final draft submitted to IPCC Panel for adoption/acceptance  
+2nd half 2027  Publication  Electronic means  
+ 
+ 
+  
+
+ 
+25  
+Decision IPCC -LXI-8.  Approval of the D raft report of the Sixtieth Session of the IPCC  
+Document: IPCC-L XI/Doc.  11, R ev.2 
+ 
+The Intergovernmental Panel on Climate Change at its Sixty -first Session approves the report of the 
+Sixtieth Session of the IPCC , as contained in document IPCC -LXI/Doc. 11, Rev.2.   
+ 
+  
+Decision IPCC -LXI- 9. Strategic Planning Schedule for the seventh assessment cycle  
+ 
+Documents: IPCC-L XI/Doc. 10; IPCC -LXI/INF. 15  
+ 
+The Intergovernmental Panel on Climate Change at its Sixt y-first Session:   
+ 
+(1) Notes the document  IPCC -LXI/Doc.10 submitted by the IPCC Chair  and document  IPCC-
+LXI/INF .15 prepared by the Co- Chairs of the Working Groups and TFI.  
+ 
+(2) Recalling  the Decision IPCC -LX-9 and in accordance with paragraph 4.1 of Appendix A of the 
+Principles  governing the work of the IPCC, based on the report s of the scoping meetings of the 
+Working Group and Task Force on National Greenhouse Gas Inventories reports, the Panel will 
+agree at its Sixty -second  Session on the scope, outline, and the work plan including schedule 
+and budget .  
+ 
+(3) Notes the Decision IPCC- LXI-5. Seventh assessment report (AR7) products – Outline of the 
+Special Report on Climate Change and Cities and Decision IPCC -LXI-7 Seventh assessment 
+report (AR7) products – Outline of the 2027 IPCC Methodology Report on Inventories for Short -
+Lived Climate Forcers.  
+   
+Decision IPCC -LXI-10. 
+ Conflict of Interest Committee on the Conflict of Interest disclosure 
+form  
+ 
+Document: IPCC- LXI/Doc. 5 
+ 
+The Intergovernmental Panel on Climate Change at its Sixty -first Session accepts the recommendations 
+of the sub-committee of the COI Committee  on the revision  of the COI disclosure form as set out in 
+Annex  I to this decision.   
+ 
+ 
+  
+ 
+    
+  
+
+ 
+26  
+ 
+ANNEX  I 
+ 
+ANNEX A 
+ 
+DRAFT REVISED COI DISCLOSURE FORM  
+ 
+CONFIDENTIAL  
+ 
+NAME:  
+ADDRESS:  
+E-MAIL ADDRESS:  
+TELEPHONE:  
+CURRENT EMPLOYER:  
+FUNCTION/ROLE IN IPCC:  
+ 
+PLEASE CONSULT THE ATTACHED GUIDANCE INFORMATION (SEE ANNEX 1) BEFORE 
+COMPLETING THE FORM BELOW  
+PLEASE FURTHER NOTE:  
+ 
+“Yes” responses do not necessarily affect or prevent your participation in IPCC activities. Answering 
+“Yes” to a question on this form does not necessarily mean that a conflict is present or that you will be unable to perform your designated function/role in the IPCC. If in doubt about whether an interest should 
+be disclosed, individuals are encouraged to disclose that information.  
+ 
+1. APPOINTMENTS AND ACTIVITY 
+ 
+ 
+Do you hold any position or appointment, or any business or professional relationships (whether 
+commercial or non- financial) with other bodies related to climate science, such as the UNFCCC or 
+others?    
+ Yes No 
+Details:  
+
+ 
+27  
+2. EMPLOYMENT AND CONSULTING  
+ 
+Do you receive any remuneration from employment or consulting, including services as a technical or 
+other adviser from a commercial entity or other organization with an interest related to the subject of the 
+IPCC work in which you are engaged?  
+ Yes     No    
+ 
+Details:  
+ 
+ 
+3. RESEARCH SUPPORT  
+ 
+Do you receive financial support (including but not limited to grants, consultancies, sponsorship, or 
+honoraria for speaking or facilitating training) or non- financial support (including but not limited to 
+premises, equipment, facilities, assistants, paid travel) from any commercial entity or other organization 
+with an interest related to the subject of the IPCC work?  
+ Yes No 
+ 
+Details:  
+ 
+ 
+4. INVESTMENT INTERESTS  
+ 
+Do you have investments (including but not limited to stocks, bonds, stock options, other securities such 
+as short sales) or commercial business interests (including but not limited to ownership, partnership, 
+joint ventures, board memberships, controlling interests), in any commercial entity with an interest 
+related to the subject of the IPCC work? (Please also i nclude indirect investments such as a trust or 
+holding company. You may exclude mutual funds, pension funds or similar investments that are broadly 
+diversified and over which you exercise no control.)  
+ Yes No 
+ 
+Details:  
+ 
+ 
+ 
+ 
+ 
+ 
+
+ 
+28 5. INTELLECTUAL PROPERTY  
+ 
+Do you own any intellectual property rights (including but not limited to patents, trademarks or 
+commercial copyrights including pending applications) or proprietary knowledge in a technology or 
+process being used for commercial purposes that might be affected by the IPCC work?  
+ Yes  No 
+ 
+Details:  
+ 
+ 
+6. PUBLIC STATEMENTS AND POSITIONS  
+ 
+As part of a regulatory, legislative or judicial process, are you providing any expert opinion or testimony 
+related to the subject of the IPCC work for a commercial entity or other organization? Yes No 
+Details:  
+ 
+ 
+7. NON- FINANCIAL INTERESTS  
+ 
+Are you engaged in any professional or other activities (including but not limited to editorial functions,  
+official (paid or unpaid) function in a government agency or international organization, advisory 
+committee associated with a public or private sector organization, board member of a public or private 
+sector organization, board member of non- profit organization, board member of advocacy group),  which 
+outside parties could consider might represent or give rise to a conflict of interest, or the perception of 
+a conflict of interest with regard the  IPCC work with which you are engaged? Yes No 
+ 
+Details:  
+ 
+ 
+8. FINANCIAL INTERESTS  
+ 
+Do you hold any additional financial interests which outside parties could consider might represent or 
+give rise to a conflict of interest, or the perception of a conflict of interest with regard to the IPCC work 
+with which you are engaged? Yes No 
+ 
+Details:  
+ 
+ 
+
+ 
+29 9. ADDITIONAL INFORMATION  
+ 
+If not already disclosed above, are you aware of any aspect of your work for the IPCC that will enable 
+you to obtain access to proprietary information or create for you a competitive advantage in your 
+professional, financial or business dealings? 
+ Yes No 
+Details:  
+ 
+ 
+To your knowledge, could the outcome of your work for the IPCC adversely affect the interests of any 
+other persons or entities with whom you have substantial common personal, professional, financial or 
+business interests (such as your adult children or siblings, close professional colleagues, administrative unit or department)?   
+ Yes No 
+ 
+Details:  
+ Which organisation is covering, partly or in full, your IPCC related travel costs?   
+ Details:  
+ 
+ Are you receiving any payments (other than for travel costs) or honoraria for speaking publicly on the 
+subject of the IPCC work in which you are engaged?  
+ Yes No 
+ 
+Details:  
+ 
+Is there any other aspect of your background or present circumstances not addressed above that you 
+consider might be perceived as affecting your objectivity or independence?                                                            
+  
+ Yes No 
+Details:  
+ 
+ 
+  
+
+ 
+30 DECLARATION  
+ 
+I hereby declare that the information in and accompanying this disclosure is true and complete to the 
+best of my knowledge and belief.  I declare that I have disclosed all associations required for disclosure 
+under the IPCC Conflict of Interest Policy; and that, except as declared, I do not consider that any of 
+the associations present a conflict of interest.  
+ 
+Should there be any change to the above information and declaration, I will promptly notify the 
+IPCC Secretariat and complete a new declaration of interest form that describes the changes. This includes any change that occurs before or during my work with the IPCC and through the period 
+of my engagement up to finalization or publication of results, or completion of the activity concerned.  
+ 
+I understand that information about my interests will be held by the IPCC for a period of five years after 
+the end of the assessment cycle during which I contributed, after which the information will be destroyed. Subject to requirement to notify the exist ence of a conflict of interest to others under paragraph 6 of the 
+Implementation Procedures, I understand that these forms will be considered confidential and will be 
+reviewed in accordance with the COI Implementation Procedures.  
+ 
+ 
+I hereby declare that I will comply with the IPCC COI Policy and the Implementation Procedures.  
+ 
+ 
+Name:  
+  
+Signature:                                                                               
+ 
+ 
+ 
+Date:  
+ 
+  
+
+ 
+31 ANNEX 1  
+ 
+GUIDANCE NOTE FOR COMPLETION OF THE CONFLICT OF INTEREST DISCLOSURE FORM  
+ 
+You have been invited to serve on the IPCC because of your professional standing and expertise. As 
+outlined in the IPCC Conflict of Interest Policy, the role of the IPCC demands that it pay special attention 
+to issues of independence and potential bias in order to maintain the integrity of, and public confidence 
+in, its products and processes. It is essential that the work of the IPCC is not compromised by any 
+conflict of interest for those who execute it. In view of this, disclosure of certain circumstances is 
+necessary to ensure that the work of the IPCC is not c ompromised by conflicts of interest. In filling out 
+this form, therefore, we rely on your professionalism, common sense, and honesty.  
+These arrangements and disclosure of interests are required as a matter of due diligence, to ensure appropriate assurance for the IPCC in matters of conflict of interest, professional and scientific integrity, 
+and to protect the IPCC and participants from reputational risk.  
+This declaration of interests, and disclosure of conflicts of interest or potential conflicts of interest, is 
+required under the IPCC Conflict of Interest Policy and Implementation Procedures.  
+You should disclose interests that could: i) significantly impair your objectivity in carrying out 
+your duties and responsibilities for the IPCC, or ii) create an unfair advantage for you or any 
+person or organization; and which could result in your securi ng a direct and material gain 
+through outcomes in an IPCC product. For the purposes of this policy, circumstances that could 
+lead a reasonable person to question your objectivity, or whether an unfair advantage has been 
+created, constitute a potential conf lict of interest and should be disclosed in this form.  
+You must also declare any relevant interests of parties with whom you have current contractual 
+relationships or substantial common interests and which could be perceived as unduly influencing, or likely to unduly influence, your judgement (for example your  employer(s), close 
+professional associates, your administrative unit or department, sponsoring or funding entities).  
+A brief  description of details should be provided in relation to any question below. You should aim to 
+provide sufficient and explicit information to allow the IPCC to form a view on whether the circumstances disclosed give rise to an actual or potential conflict  of interest. If in doubt about whether an interest 
+should be disclosed, individuals are encouraged to disclose that information.  
+Please sign and date this form on the last page, and return the form to the Secretary of the IPCC with 
+a Curriculum Vitae  and information supporting these disclosures where applicable.   Retain a copy for 
+your records.  
+You must promptly inform the IPCC Secretariat if there is any change in this information prior to or during the course of your work or meetings for the IPCC. This form and the declarations contained 
+therein must be completed before participation in the IPC C activity can be confirmed.  
+Answering “Yes” to a question on this form does not necessarily mean that a conflict is present 
+or that you will be unable to perform your designated function/role in the IPCC. If in doubt about whether an interest should be disclosed, individuals are encouraged to disclose that information. This information will be assessed as a whole on the basis of the principles 
+contained in the COI Policy (https://www.ipcc.ch/site/assets/uploads/2018/09/ipcc -conflict -of-
+interest -2016.pdf). In particular, what constitut es or not a COI is defined in paragraphs 11 to 17 
+of that document (reproduced below).  If in doubt about whether an interest should be disclosed,  
+
+ 
+32  
+individuals are encouraged to seek advice from IPCC Secretariat Legal Officer (please contact ipcc-
+[email protected]  f or contact information).  
+Definition of « Conflict of Interest  » (paragraphs 11 to 17 of the IPCC COI Policy 
+https://www.ipcc.ch/site/assets/uploads/2018/09/ipcc -conflict -of-interest -2016.pdf).  
+Conflict of Interest  
+11. A “conflict of interest” refers to any current professional, financial or other interest which could: i) 
+significantly impair the individual’s objectivity in carrying out his or her duties and responsibilities for the 
+IPCC, or ii) create an unfair advantage for any person or organization. For the purposes of this policy, circumstances that could lead a reasonable person to question an individual’s objectivity, or whether an unfair advantage has been created, constitute a potential conflict of interest. These potential conflicts 
+are subject to disclosure.  
+ 
+12. Conflict of interest policies in scientific assessment bodies typically make a distinction between “conflict of interest” and “bias,” which refers to a point of view or perspective that is strongly held 
+regarding a particular issue or set of issues. In the case of author and review teams, bias can and should 
+be managed through the selection of a balance of perspectives. For example, it is expected that IPCC 
+author teams will include individuals with different perspectives and affiliations. Those involved in selecting authors will need to strive for an author team composition that reflects a balance of expertise 
+and perspectives, such that IPCC products are comprehensive, objective, and  
+neutral with respect to policy. In selecting these individuals, care must be taken to ensure that biases 
+can be balanced where they exist. In contrast, conflict of interest exists where an individual could secure 
+a direct and material gain through outcomes  in an IPCC product. Holding a view that one believes to be 
+correct, but that one does not stand to gain from personally is not a conflict of interest.  
+ 
+13. The conflict of interest requirements in this policy are not designed to include an assessment of 
+one's behavior or character or one's ability to act objectively despite the conflict of interest.  
+ 
+14. This policy applies only to current conflicts of interest. It does not apply to past interests that have expired, no longer exist, and cannot reasonably affect current behavior. Nor does it apply to possible 
+interests that may arise in the future but that do not currently exist, as such interests are inherently speculative and uncertain. For example, a pending application for a particular job is a current interest, but the mere possibility that one mi ght apply for such a job in the future is not a current interest.  
+ 
+15. Professional and other non -financial interests need to be disclosed only if they are significant and 
+relevant. If in doubt about whether an interest should be disclosed, individuals are encouraged to seek 
+advice from the appropriate IPCC body as defined in Annex A. Significant and relevant interests may include, but are not limited to, senior editorial roles, advisory committees associated with private sector 
+organizations, and memberships on boards of non- profit or advocacy groups. However, not all suc h 
+associations necessarily constitute a conflict of interest. 
+ 16. Financial interests need to be disclosed only if they are significant and relevant. These may include, 
+but are not limited to, the following kinds of financial interests: employment relationships; consulting 
+relationships; financial investments; intellectual property interests; and commercial interests and sources of private- sector research support. Individuals should also disclose significant  and relevant 
+financial interests of any person with whom the individual has a substantial business or  relevant shared 
+interest. If in doubt about whether an interest should be disclosed, individuals are encouraged to seek 
+advice from the appropriate IPCC body as defined in Annex A “Implementation”.  
+
+ 
+33  
+17. To prevent situations in which a conflict of interest may arise, individuals directly involved in or leading the preparation of IPCC reports should avoid being in a position to approve, adopt, or accept 
+on behalf of any government the text in which he/she was directly involved.  
+  
+
+ 
+34  
+Decision IPCC -LXI-11. Matters related to other IPCC activities – Terms of Reference of the IPCC 
+Publications Committee 
+ 
+Document: IPCC-L XI/Doc. 4 
+ 
+The Intergovernmental Panel on Climate Change at its Sixt y-first Session agrees  on the Terms of 
+Reference of the IPCC Publication Committee , as contained in Annex 1 to this decision.  
+  
+
+ 
+35 ANNEX  1 
+ 
+ 
+DRAFT TERMS OF REFERENCE OF THE IPCC PUBLICATIONS COMMITTEE  
+ 
+ 
+Terms of Reference  
+ 
+1. The IPCC Publications and Translations Committee (hereafter known as the “Committee”) Terms 
+of Reference are intended to be in line with  and not conflict with the IPCC principles and procedures.  
+ 
+Purpose and Scope  
+ 
+2. The Committee is established for the duration of the respective assessment cycle, to oversee the 
+implementation of the recommendations of the Panel and Bureau with regards to publications, 
+translations and access to literature and advise the IPCC Secretariat on:  
+ 
+a. Technical specifications and Terms of Reference for procurement of WMO translation 
+services;  
+b. Technical specifications and Terms of Reference for procurement processes for printing and 
+publishing services for IPCC products;  
+c. Management of citation data for past and future IPCC reports and their main components;   
+d. Timely establishment of editorial sub- committees for translation into each official UN 
+language;  
+e. Proposals for enhancing quality and review of translations of scientific and technical IPCC products;  
+f. Options for enhancing access to literature for IPCC authors.  
+ 
+Appointment of Members 
+ 3. The Committee shall be composed of nine members:  
+ 
+• two from each Working Group and Task Force for Inventories Bureau;  
+• one IPCC Vice Chair to be the Chair of the Committee.  
+ 
+Additionally, the Head of the IPCC Secretariat and Co- Chairs of TG -Data, or their delegates, will 
+serve in an advisory role to the Committee.  
+ 4. The members will be appointed by their respective Working Group and Task Force Co -Chairs taking 
+into account overall gender and regional representation, with a view to collective UN language 
+expertise. The Chair to the Committee will be appointed by the IPCC Chair from amongst the IPCC 
+Vice-Chairs.  
+ 
+5. Working Group and TFI members will be supported by their respective TSUs, as needed.  
+ 
+Modus operandi  
+ 6. The Committee:  
+ 
+a. Will meet as necessary at a time and location to be established by the Chair of the 
+Committee. Such meetings may take place by electronic means unless they are organized in 
+the margins of other IPCC meetings which will take place in person;  
+b. Will reach decisions by consensus; where consensus is deemed not possible, the matter will 
+be referred back to the Bureau;  
+
+ 
+36 c. Five members of the Committee including the IPCC Vice- Chair shall constitute a quorum;  
+d. Will liaise with the WMO Publication Board to ensure coordination, planning and scheduling 
+related to establishment of a WMO Tender Evaluation Board (“TEB”) and in the 
+bidding/evaluation process for IPCC publications and any related products;  
+e. Have at least two members of the Committee offer to serve on the TEB for an IPCC 
+publication/procurement process overseen by the WMO. Such Committee members will 
+serve on the TEB in their personal capacity and will need to be able to meet the neutrality 
+and conflict of interest test for membership;  
+ may seek advice from qualified experts, such as librarians, publishing organizations and 
+international scientific bodies;  
+f. [Will identify options for the expansion of access to literature for IPCC authors and for 
+implementing these following guidance of the IPCC Bureau ; 
+g. Will undertake to prepare best practices for producing translations of IPCC products;  
+h. Will agree annually on an implementation plan; 
+i. Will report regularly to the Bureau.  
+ 
+   
+ 
+ 
+

IPCC_AR6_WGIII_FullReport.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24c16754a7dff6e03147fe74825794240eccc970a36e5e7f7929a4afb7531ec5
+size 12042003

SPM_version_report_LR.txt

@@ -0,0 +1,1164 @@
+Summary for Policymakers
+
+SPM3SPM
+Drafting Authors:
+Myles Allen (UK), Mustafa Babiker (Sudan), Yang Chen (China), Heleen de Coninck 
+(Netherlands/EU), Sarah Connors (UK), Renée van Diemen (Netherlands), Opha Pauline 
+Dube (Botswana), Kristie L. Ebi (USA), Francois Engelbrecht (South Africa), Marion Ferrat 
+(UK/France), James Ford (UK/Canada), Piers Forster (UK), Sabine Fuss (Germany), Tania 
+Guillén Bolaños (Germany/Nicaragua), Jordan Harold (UK), Ove Hoegh-Guldberg (Australia), 
+Jean-Charles Hourcade (France), Daniel Huppmann (Austria), Daniela Jacob (Germany), 
+Kejun Jiang (China), Tom Gabriel Johansen (Norway), Mikiko Kainuma (Japan), Kiane de 
+Kleijne (Netherlands/EU), Elmar Kriegler (Germany), Debora Ley (Guatemala/Mexico), 
+Diana Liverman (USA), Natalie Mahowald (USA), Valérie Masson-Delmotte (France),  
+J. B. Robin Matthews (UK), Richard Millar (UK), Katja Mintenbeck (Germany), Angela Morelli 
+(Norway/Italy), Wilfran Moufouma-Okia (France/Congo), Luis Mundaca (Sweden/Chile), 
+Maike Nicolai (Germany), Chukwumerije Okereke (UK/Nigeria), Minal Pathak (India), Antony 
+Payne (UK), Roz Pidcock (UK), Anna Pirani (Italy), Elvira Poloczanska (UK/Australia), Hans-
+Otto Pörtner (Germany), Aromar Revi (India), Keywan Riahi (Austria), Debra C. Roberts 
+(South Africa), Joeri Rogelj (Austria/Belgium), Joyashree Roy (India), Sonia I. Seneviratne 
+(Switzerland), Priyadarshi R. Shukla (India), James Skea (UK), Raphael Slade (UK), Drew 
+Shindell (USA), Chandni Singh (India), William Solecki (USA), Linda Steg (Netherlands), 
+Michael Taylor (Jamaica), Petra Tschakert (Australia/Austria), Henri Waisman (France), 
+Rachel Warren (UK), Panmao Zhai (China), Kirsten Zickfeld (Canada).
+This Summary for Policymakers should be cited as:
+IPCC, 2018: Summary for Policymakers. In: Global Warming of 1.5°C. An IPCC Special Report on the impacts 
+of global warming of 1.5°C above pre-industrial levels and related global greenhouse gas emission pathways, 
+in the context of strengthening the global response to the threat of climate change, sustainable development, 
+and efforts to eradicate poverty [Masson-Delmotte, V., P . Zhai, H.-O.  Pörtner, D.  Roberts, J. Skea, P .R. Shukla, 
+A. Pirani, W. Moufouma-Okia, C. Péan, R. Pidcock, S. Connors, J.B.R. Matthews, Y . Chen, X.  Zhou, M.I.  Gomis, 
+E. Lonnoy, T. Maycock, M. Tignor, and T. Waterfield (eds.)]. Cambridge University Press, Cambridge, UK and New 
+York, NY , USA, pp. 3-24. https://doi.org/10.1017/9781009157940.001.Summary 
+for Policymakers SPM
+
+SPMSummary for Policymakers4Introduction
+This Report responds to the invitation for IPCC ‘... to provide a Special Report in 2018 on the impacts of global warming of 1.5°C 
+above pre-industrial levels and related global greenhouse gas emission pathways’ contained in the Decision of the 21st Conference 
+of Parties of the United Nations Framework Convention on Climate Change to adopt the Paris Agreement.1
+The IPCC accepted the invitation in April 2016, deciding to prepare this Special Report on the impacts of global warming of 
+1.5°C above pre-industrial levels and related global greenhouse gas emission pathways, in the context of strengthening the global 
+response to the threat of climate change, sustainable development, and efforts to eradicate poverty.
+This Summary for Policymakers (SPM) presents the key findings of the Special Report, based on the assessment of the available 
+scientific, technical and socio-economic literature2 relevant to global warming of 1.5°C and for the comparison between global 
+warming of 1.5°C and 2°C above pre-industrial levels. The level of confidence associated with each key finding is reported using 
+the IPCC calibrated language.3 The underlying scientific basis of each key finding is indicated by references provided to chapter 
+elements. In the SPM, knowledge gaps are identified associated with the underlying chapters of the Report.
+A. Understanding Global Warming of 1.5°C4
+A.1 Human activities are estimated to have caused approximately 1.0°C of global warming5 above 
+pre-industrial levels, with a likely range of 0.8°C to 1.2°C. Global warming is likely to reach 1.5°C 
+between 2030 and 2052 if it continues to increase at the current rate. (high confidence) (Figure 
+SPM.1) {1.2}
+A.1.1 Reflecting the long-term warming trend since pre-industrial times, observed global mean surface temperature (GMST) for 
+the decade 2006–2015 was 0.87°C (likely between 0.75°C and 0.99°C)6 higher than the average over the 1850–1900 
+period (very high confidence). Estimated anthropogenic global warming matches the level of observed warming to within 
+±20% (likely range). Estimated anthropogenic global warming is currently increasing at 0.2°C (likely between 0.1°C and 
+0.3°C) per decade due to past and ongoing emissions (high confidence). {1.2.1, Table 1.1, 1.2.4}
+A.1.2 Warming greater than the global annual average is being experienced in many land regions and seasons, including two to 
+three times higher in the Arctic. Warming is generally higher over land than over the ocean. (high confidence) {1.2.1, 1.2.2, 
+Figure 1.1, Figure 1.3, 3.3.1, 3.3.2}
+A.1.3 Trends in intensity and frequency of some climate and weather extremes have been detected over time spans during which 
+about 0.5°C of global warming occurred (medium confidence). This assessment is based on several lines of evidence, 
+including attribution studies for changes in extremes since 1950. {3.3.1, 3.3.2, 3.3.3} 
+1 Decision 1/CP .21, paragraph 21.
+2 The assessment covers literature accepted for publication by 15 May 2018.
+3 Each finding is grounded in an evaluation of underlying evidence and agreement. A level of confidence is expressed using five qualifiers: very low, low, medium, high and very high, and  
+ typeset in italics, for example, medium confidence. The following terms have been used to indicate the assessed likelihood of an outcome or a result: virtually certain 99–100%  
+ probability, very likely 90–100%, likely 66–100%, about as likely as not 33–66%, unlikely 0–33%, very unlikely 0–10%, exceptionally unlikely 0–1%. Additional terms (extremely likely  
+ 95–100%, more likely than not >50–100%, more unlikely than likely 0–<50%, extremely unlikely 0–5%) may also be used when appropriate. Assessed likelihood is typeset in italics,  
+ for example, very likely. This is consistent with AR5. 
+4 See also Box SPM.1: Core Concepts Central to this Special Report.
+5 Present level of global warming is defined as the average of a 30-year period centred on 2017 assuming the recent rate of warming continues.
+6 This range spans the four available peer-reviewed estimates of the observed GMST change and also accounts for additional uncertainty due to possible short-term natural variability.  
+ {1.2.1, Table 1.1}
+
+SPM Summary for Policymakers5A.2 Warming from anthropogenic emissions from the pre-industrial period to the present will persist for 
+centuries to millennia and will continue to cause further long-term changes in the climate system, 
+such as sea level rise, with associated impacts (high confidence), but these emissions alone are 
+unlikely to cause global warming of 1.5°C (medium confidence). (Figure SPM.1) {1.2, 3.3, Figure 1.5}
+A.2.1 Anthropogenic emissions (including greenhouse gases, aerosols and their precursors) up to the present are unlikely to 
+cause further warming of more than 0.5°C over the next two to three decades (high confidence) or on a century time scale 
+(medium confidence). {1.2.4, Figure 1.5}
+A.2.2 Reaching and sustaining net zero global anthropogenic CO2 emissions and declining net non-CO2 radiative forcing would 
+halt anthropogenic global warming on multi-decadal time scales (high confidence). The maximum temperature reached is 
+then determined by cumulative net global anthropogenic CO2 emissions up to the time of net zero CO2 emissions (high 
+confidence) and the level of non-CO2 radiative forcing in the decades prior to the time that maximum temperatures are 
+reached (medium confidence). On longer time scales, sustained net negative global anthropogenic CO2 emissions and/
+or further reductions in non-CO2 radiative forcing may still be required to prevent further warming due to Earth system 
+feedbacks and to reverse ocean acidification (medium confidence) and will be required to minimize sea level rise (high 
+confidence). {Cross-Chapter Box 2 in Chapter 1, 1.2.3, 1.2.4, Figure 1.4, 2.2.1, 2.2.2, 3.4.4.8, 3.4.5.1, 3.6.3.2}
+A.3 Climate-related risks for natural and human systems are higher for global warming of 1.5°C than 
+at present, but lower than at 2°C (high confidence). These risks depend on the magnitude and rate 
+of warming, geographic location, levels of development and vulnerability, and on the choices and 
+implementation of adaptation and mitigation options (high confidence). (Figure SPM.2) {1.3, 3.3, 
+3.4, 5.6}
+A.3.1 Impacts on natural and human systems from global warming have already been observed (high confidence). Many land and 
+ocean ecosystems and some of the services they provide have already changed due to global warming (high confidence). 
+(Figure SPM.2) {1.4, 3.4, 3.5}
+A.3.2 Future climate-related risks depend on the rate, peak and duration of warming. In the aggregate, they are larger if global 
+warming exceeds 1.5°C before returning to that level by 2100 than if global warming gradually stabilizes at 1.5°C, especially 
+if the peak temperature is high (e.g., about 2°C) (high confidence). Some impacts may be long-lasting or irreversible, such 
+as the loss of some ecosystems (high confidence). {3.2, 3.4.4, 3.6.3, Cross-Chapter Box 8 in Chapter 3}
+A.3.3 Adaptation and mitigation are already occurring (high confidence). Future climate-related risks would be reduced by the 
+upscaling and acceleration of far-reaching, multilevel and cross-sectoral climate mitigation and by both incremental and 
+transformational adaptation (high confidence). {1.2, 1.3, Table 3.5, 4.2.2, Cross-Chapter Box 9 in Chapter 4, Box 4.2, Box 
+4.3, Box 4.6, 4.3.1, 4.3.2, 4.3.3, 4.3.4, 4.3.5, 4.4.1, 4.4.4, 4.4.5, 4.5.3}  
+
+SPMSummary for Policymakers660
+503 000
+2 000
+1 00040
+30
+20
+10
+0 03
+2
+1
+0Cumulative emissions of CO/two.dnom and future non-CO/two.dnom radiative forcing determine 
+the probability of limiting warming to 1.5°C
+Billion tonnes CO/two.dnom per year (GtCO/two.dnom/yr) Billion tonnes CO/two.dnom (GtCO/two.dnom) Watts per square metre (W/m/two.numr)b) Stylized net global CO/two.dnom emission pathways d) Non-CO/two.dnom radiative forcing pathways c) Cumulative net CO/two.dnom emissionsa) Observed global temperature change and modeled 
+responses to stylized anthropogenic emission and forcing pathways
+Observed monthly global 
+mean surface temperature
+Estimated anthropogenic 
+warming to date and 
+likely range
+Faster immediate CO/two.dnom emission reductions 
+limit cumulative CO/two.dnom emissions shown in 
+panel (c).Maximum temperature rise is determined by cumulative net CO/two.dnom emissions and net non-CO/two.dnom 
+radiative forcing due to methane, nitrous oxide, aerosols and other anthropogenic forcing agents.Global warming relative to 1850-1900 (°C)
+Cumulative CO/two.dnom 
+emissions in pathways 
+reaching net zero in 
+2055 and 2040Non-CO/two.dnom radiative forcing 
+reduced a/f_ter 2030 or 
+not reduced a/f_ter 20301960
+1980 2020 2060 2100 1980 2020 2060 2100 1980 2020 2060 21001980 2000 20202017
+2040 2060 2080 21002.0
+1.5
+1.0
+0.5
+0
+Likely range of modeled responses to stylized pathways
+      Faster CO/two.dnom reductions (blue in b & c) result in a higher 
+probability of limiting warming to 1.5°C 
+      No reduction of net non-CO/two.dnom radiative forcing (purple in d) 
+results in a lower probability of limiting warming to 1.5°C       Global CO/two.dnom emissions reach net zero in 2055 while net 
+non-CO/two.dnom radiative forcing is reduced a/f_ter 2030 (grey in b, c & d)
+Figure SPM.1 | Panel a: Observed monthly global mean surface temperature (GMST, grey line up to 2017, from the HadCRUT4, GISTEMP , Cowtan–Way, and 
+NOAA datasets) change and estimated anthropogenic global warming (solid orange line up to 2017, with orange shading indicating assessed likely range). Orange 
+dashed arrow and horizontal orange error bar show respectively the central estimate and likely range of the time at which 1.5°C is reached if the current rate 
+of warming continues. The grey plume on the right of panel a shows the likely range of warming responses, computed with a simple climate model, to a stylized 
+pathway (hypothetical future) in which net CO2 emissions (grey line in panels b and c) decline in a straight line from 2020 to reach net zero in 2055 and net non-
+CO2 radiative forcing (grey line in panel d) increases to 2030 and then declines. The blue plume in panel a) shows the response to faster CO2 emissions reductions 
+(blue line in panel b), reaching net zero in 2040, reducing cumulative CO2 emissions (panel c). The purple plume shows the response to net CO2 emissions declining 
+to zero in 2055, with net non-CO2 forcing remaining constant after 2030. The vertical error bars on right of panel a) show the likely ranges (thin lines) and central 
+terciles (33rd – 66th percentiles, thick lines) of the estimated distribution of warming in 2100 under these three stylized pathways. Vertical dotted error bars in 
+panels b, c and d show the likely range of historical annual and cumulative global net CO2 emissions in 2017 (data from the Global Carbon Project) and of net 
+non-CO2 radiative forcing in 2011 from AR5, respectively. Vertical axes in panels c and d are scaled to represent approximately equal effects on GMST. {1.2.1, 1.2.3, 
+1.2.4, 2.3, Figure 1.2 and Chapter 1 Supplementary Material, Cross-Chapter Box 2 in Chapter 1}
+
+SPM Summary for Policymakers7B. Projected Climate Change, Potential Impacts and Associated Risks
+B.1 Climate models project robust7 differences in regional climate characteristics between present-day 
+and global warming of 1.5°C,8 and between 1.5°C and 2°C.8 These differences include increases 
+in: mean temperature in most land and ocean regions (high confidence), hot extremes in most 
+inhabited regions (high confidence), heavy precipitation in several regions (medium confidence), 
+and the probability of drought and precipitation deficits in some regions (medium confidence). 
+{3.3}
+B.1.1 Evidence from attributed changes in some climate and weather extremes for a global warming of about 0.5°C supports 
+the assessment that an additional 0.5°C of warming compared to present is associated with further detectable changes in 
+these extremes (medium confidence). Several regional changes in climate are assessed to occur with global warming up 
+to 1.5°C compared to pre-industrial levels, including warming of extreme temperatures in many regions (high confidence), 
+increases in frequency, intensity, and/or amount of heavy precipitation in several regions (high confidence), and an increase 
+in intensity or frequency of droughts in some regions (medium confidence). {3.2, 3.3.1, 3.3.2, 3.3.3, 3.3.4, Table 3.2}
+B.1.2 Temperature extremes on land are projected to warm more than GMST (high confidence): extreme hot days in mid-latitudes 
+warm by up to about 3°C at global warming of 1.5°C and about 4°C at 2°C, and extreme cold nights in high latitudes warm 
+by up to about 4.5°C at 1.5°C and about 6°C at 2°C (high confidence). The number of hot days is projected to increase in 
+most land regions, with highest increases in the tropics (high confidence). {3.3.1, 3.3.2, Cross-Chapter Box 8 in Chapter 3}
+B.1.3 Risks from droughts and precipitation deficits are projected to be higher at 2°C compared to 1.5°C of global warming in 
+some regions (medium confidence). Risks from heavy precipitation events are projected to be higher at 2°C compared to 
+1.5°C of global warming in several northern hemisphere high-latitude and/or high-elevation regions, eastern Asia and 
+eastern North America (medium confidence). Heavy precipitation associated with tropical cyclones is projected to be 
+higher at 2°C compared to 1.5°C global warming (medium confidence). There is generally low confidence in projected 
+changes in heavy precipitation at 2°C compared to 1.5°C in other regions. Heavy precipitation when aggregated at global 
+scale is projected to be higher at 2°C than at 1.5°C of global warming (medium confidence). As a consequence of heavy 
+precipitation, the fraction of the global land area affected by flood hazards is projected to be larger at 2°C compared to 
+1.5°C of global warming (medium confidence). {3.3.1, 3.3.3, 3.3.4, 3.3.5, 3.3.6}
+B.2 By 2100, global mean sea level rise is projected to be around 0.1 metre lower with global warming 
+of 1.5°C compared to 2°C (medium confidence). Sea level will continue to rise well beyond 2100 
+(high confidence), and the magnitude and rate of this rise depend on future emission pathways. 
+A slower rate of sea level rise enables greater opportunities for adaptation in the human and 
+ecological systems of small islands, low-lying coastal areas and deltas (medium confidence).  
+{3.3, 3.4, 3.6}
+B.2.1 Model-based projections of global mean sea level rise (relative to 1986–2005) suggest an indicative range of 0.26 to 0.77 
+m by 2100 for 1.5°C of global warming, 0.1 m (0.04–0.16 m) less than for a global warming of 2°C (medium confidence). 
+A reduction of 0.1 m in global sea level rise implies that up to 10 million fewer people would be exposed to related risks, 
+based on population in the year 2010 and assuming no adaptation (medium confidence). {3.4.4, 3.4.5, 4.3.2}
+B.2.2 Sea level rise will continue beyond 2100 even if global warming is limited to 1.5°C in the 21st century (high confidence). 
+Marine ice sheet instability in Antarctica and/or irreversible loss of the Greenland ice sheet could result in multi-metre rise 
+in sea level over hundreds to thousands of years. These instabilities could be triggered at around 1.5°C to 2°C of global 
+warming (medium confidence). (Figure SPM.2) {3.3.9, 3.4.5, 3.5.2, 3.6.3, Box 3.3}
+7 Robust is here used to mean that at least two thirds of climate models show the same sign of changes at the grid point scale, and that differences in large regions are statistically  
+ significant.
+8 Projected changes in impacts between different levels of global warming are determined with respect to changes in global mean surface air temperature.
+
+SPMSummary for Policymakers8B.2.3 Increasing warming amplifies the exposure of small islands, low-lying coastal areas and deltas to the risks associated with 
+sea level rise for many human and ecological systems, including increased saltwater intrusion, flooding and damage to 
+infrastructure (high confidence). Risks associated with sea level rise are higher at 2°C compared to 1.5°C. The slower rate 
+of sea level rise at global warming of 1.5°C reduces these risks, enabling greater opportunities for adaptation including 
+managing and restoring natural coastal ecosystems and infrastructure reinforcement (medium confidence). (Figure SPM.2) 
+{3.4.5, Box 3.5}
+B.3 On land, impacts on biodiversity and ecosystems, including species loss and extinction, are 
+projected to be lower at 1.5°C of global warming compared to 2°C. Limiting global warming to 
+1.5°C compared to 2°C is projected to lower the impacts on terrestrial, freshwater and coastal 
+ecosystems and to retain more of their services to humans (high confidence). (Figure SPM.2)  
+{3.4, 3.5, Box 3.4, Box 4.2, Cross-Chapter Box 8 in Chapter 3} 
+B.3.1 Of 105,000 species studied,9 6% of insects, 8% of plants and 4% of vertebrates are projected to lose over half of their 
+climatically determined geographic range for global warming of 1.5°C, compared with 18% of insects, 16% of plants and 
+8% of vertebrates for global warming of 2°C (medium confidence). Impacts associated with other biodiversity-related 
+risks such as forest fires and the spread of invasive species are lower at 1.5°C compared to 2°C of global warming (high 
+confidence). {3.4.3, 3.5.2}
+B.3.2 Approximately 4% (interquartile range 2–7%) of the global terrestrial land area is projected to undergo a transformation 
+of ecosystems from one type to another at 1°C of global warming, compared with 13% (interquartile range 8–20%) at 2°C 
+(medium confidence). This indicates that the area at risk is projected to be approximately 50% lower at 1.5°C compared to 
+2°C (medium confidence). {3.4.3.1, 3.4.3.5}
+B.3.3 High-latitude tundra and boreal forests are particularly at risk of climate change-induced degradation and loss, with woody 
+shrubs already encroaching into the tundra (high confidence) and this will proceed with further warming. Limiting global 
+warming to 1.5°C rather than 2°C is projected to prevent the thawing over centuries of a permafrost area in the range of 
+1.5 to 2.5 million km2 (medium confidence). {3.3.2, 3.4.3, 3.5.5} 
+B.4 Limiting global warming to 1.5°C compared to 2°C is projected to reduce increases in ocean 
+temperature as well as associated increases in ocean acidity and decreases in ocean oxygen levels 
+(high confidence). Consequently, limiting global warming to 1.5°C is projected to reduce risks 
+to marine biodiversity, fisheries, and ecosystems, and their functions and services to humans, 
+as illustrated by recent changes to Arctic sea ice and warm-water coral reef ecosystems (high 
+confidence). {3.3, 3.4, 3.5, Box 3.4, Box 3.5}
+B.4.1 There is high confidence that the probability of a sea ice-free Arctic Ocean during summer is substantially lower at global 
+warming of 1.5°C when compared to 2°C. With 1.5°C of global warming, one sea ice-free Arctic summer is projected per 
+century. This likelihood is increased to at least one per decade with 2°C global warming. Effects of a temperature overshoot 
+are reversible for Arctic sea ice cover on decadal time scales (high confidence). {3.3.8, 3.4.4.7}
+B.4.2  Global warming of 1.5°C is projected to shift the ranges of many marine species to higher latitudes as well as increase the 
+amount of damage to many ecosystems. It is also expected to drive the loss of coastal resources and reduce the productivity of 
+fisheries and aquaculture (especially at low latitudes). The risks of climate-induced impacts are projected to be higher at 2°C 
+than those at global warming of 1.5°C (high confidence). Coral reefs, for example, are projected to decline by a further 70–90% 
+at 1.5°C (high confidence) with larger losses (>99%) at 2°C (very high confidence). The risk of irreversible loss of many marine 
+and coastal ecosystems increases with global warming, especially at 2°C or more (high confidence). {3.4.4, Box 3.4}
+9 Consistent with earlier studies, illustrative numbers were adopted from one recent meta-study.
+
+SPM Summary for Policymakers910 Here, impacts on economic growth refer to changes in gross domestic product (GDP). Many impacts, such as loss of human lives, cultural heritage and ecosystem services, are difficult 
+to value and monetize.B.4.3  The level of ocean acidification due to increasing CO2 concentrations associated with global warming of 1.5°C is projected to 
+amplify the adverse effects of warming, and even further at 2°C, impacting the growth, development, calcification, survival, 
+and thus abundance of a broad range of species, for example, from algae to fish (high confidence). {3.3.10, 3.4.4}
+B.4.4  Impacts of climate change in the ocean are increasing risks to fisheries and aquaculture via impacts on the physiology, 
+survivorship, habitat, reproduction, disease incidence, and risk of invasive species (medium confidence) but are projected to 
+be less at 1.5°C of global warming than at 2°C. One global fishery model, for example, projected a decrease in global annual 
+catch for marine fisheries of about 1.5 million tonnes for 1.5°C of global warming compared to a loss of more than 3 million 
+tonnes for 2°C of global warming (medium confidence). {3.4.4, Box 3.4}
+B.5 Climate-related risks to health, livelihoods, food security, water supply, human security, and 
+economic growth are projected to increase with global warming of 1.5°C and increase further with 
+2°C. (Figure SPM.2) {3.4, 3.5, 5.2, Box 3.2, Box 3.3, Box 3.5, Box 3.6, Cross-Chapter Box 6 in Chapter 
+3, Cross-Chapter Box 9 in Chapter 4, Cross-Chapter Box 12 in Chapter 5, 5.2} 
+B.5.1 Populations at disproportionately higher risk of adverse consequences with global warming of 1.5°C and beyond include 
+disadvantaged and vulnerable populations, some indigenous peoples, and local communities dependent on agricultural or 
+coastal livelihoods (high confidence). Regions at disproportionately higher risk include Arctic ecosystems, dryland regions, 
+small island developing states, and Least Developed Countries (high confidence). Poverty and disadvantage are expected 
+to increase in some populations as global warming increases; limiting global warming to 1.5°C, compared with 2°C, could 
+reduce the number of people both exposed to climate-related risks and susceptible to poverty by up to several hundred 
+million by 2050 (medium confidence). {3.4.10, 3.4.11, Box 3.5, Cross-Chapter Box 6 in Chapter 3, Cross-Chapter Box 9 in 
+Chapter 4, Cross-Chapter Box 12 in Chapter 5, 4.2.2.2, 5.2.1, 5.2.2, 5.2.3, 5.6.3}
+B.5.2 Any increase in global warming is projected to affect human health, with primarily negative consequences (high confidence). 
+Lower risks are projected at 1.5°C than at 2°C for heat-related morbidity and mortality (very high confidence) and for 
+ozone-related mortality if emissions needed for ozone formation remain high (high confidence). Urban heat islands often 
+amplify the impacts of heatwaves in cities (high confidence). Risks from some vector-borne diseases, such as malaria and 
+dengue fever, are projected to increase with warming from 1.5°C to 2°C, including potential shifts in their geographic range 
+(high confidence). {3.4.7, 3.4.8, 3.5.5.8}
+B.5.3 Limiting warming to 1.5°C compared with 2°C is projected to result in smaller net reductions in yields of maize, rice, wheat, 
+and potentially other cereal crops, particularly in sub-Saharan Africa, Southeast Asia, and Central and South America, and 
+in the CO2-dependent nutritional quality of rice and wheat (high confidence). Reductions in projected food availability are 
+larger at 2°C than at 1.5°C of global warming in the Sahel, southern Africa, the Mediterranean, central Europe, and the 
+Amazon (medium confidence). Livestock are projected to be adversely affected with rising temperatures, depending on the 
+extent of changes in feed quality, spread of diseases, and water resource availability (high confidence). {3.4.6, 3.5.4, 3.5.5, 
+Box 3.1, Cross-Chapter Box 6 in Chapter 3, Cross-Chapter Box 9 in Chapter 4}
+B.5.4 Depending on future socio-economic conditions, limiting global warming to 1.5°C compared to 2°C may reduce the 
+proportion of the world population exposed to a climate change-induced increase in water stress by up to 50%, although 
+there is considerable variability between regions (medium confidence). Many small island developing states could  
+experience lower water stress as a result of projected changes in aridity when global warming is limited to 1.5°C, as 
+compared to 2°C (medium confidence). {3.3.5, 3.4.2, 3.4.8, 3.5.5, Box 3.2, Box 3.5, Cross-Chapter Box 9 in Chapter 4}
+B.5.5 Risks to global aggregated economic growth due to climate change impacts are projected to be lower at 1.5°C than at 
+2°C by the end of this century10 (medium confidence). This excludes the costs of mitigation, adaptation investments and 
+the benefits of adaptation. Countries in the tropics and Southern Hemisphere subtropics are projected to experience the 
+largest impacts on economic growth due to climate change should global warming increase from 1.5°C to 2°C (medium 
+confidence). {3.5.2, 3.5.3} 
+
+SPMSummary for Policymakers10B.5.6 Exposure to multiple and compound climate-related risks increases between 1.5°C and 2°C of global warming, with greater 
+proportions of people both so exposed and susceptible to poverty in Africa and Asia (high confidence). For global warming 
+from 1.5°C to 2°C, risks across energy, food, and water sectors could overlap spatially and temporally, creating new and 
+exacerbating current hazards, exposures, and vulnerabilities that could affect increasing numbers of people and regions 
+(medium confidence). {Box 3.5, 3.3.1, 3.4.5.3, 3.4.5.6, 3.4.11, 3.5.4.9}
+B.5.7 There are multiple lines of evidence that since AR5 the assessed levels of risk increased for four of the five Reasons for 
+Concern (RFCs) for global warming to 2°C (high confidence). The risk transitions by degrees of global warming are now: 
+from high to very high risk between 1.5°C and 2°C for RFC1 (Unique and threatened systems) (high confidence); from 
+moderate to high risk between 1°C and 1.5°C for RFC2 (Extreme weather events) (medium confidence); from moderate to 
+high risk between 1.5°C and 2°C for RFC3 (Distribution of impacts) (high confidence); from moderate to high risk between 
+1.5°C and 2.5°C for RFC4 (Global aggregate impacts) (medium confidence); and from moderate to high risk between 1°C 
+and 2.5°C for RFC5 (Large-scale singular events) (medium confidence). (Figure SPM.2) {3.4.13; 3.5, 3.5.2}
+B.6  Most adaptation needs will be lower for global warming of 1.5°C compared to 2°C (high confidence). 
+There are a wide range of adaptation options that can reduce the risks of climate change (high 
+confidence). There are limits to adaptation and adaptive capacity for some human and natural 
+systems at global warming of 1.5°C, with associated losses (medium confidence). The number and 
+availability of adaptation options vary by sector (medium confidence). {Table 3.5, 4.3, 4.5, Cross-
+Chapter Box 9 in Chapter 4, Cross-Chapter Box 12 in Chapter 5} 
+B.6.1 A wide range of adaptation options are available to reduce the risks to natural and managed ecosystems (e.g., ecosystem-
+based adaptation, ecosystem restoration and avoided degradation and deforestation, biodiversity management, 
+sustainable aquaculture, and local knowledge and indigenous knowledge), the risks of sea level rise (e.g., coastal defence 
+and hardening), and the risks to health, livelihoods, food, water, and economic growth, especially in rural landscapes 
+(e.g., efficient irrigation, social safety nets, disaster risk management, risk spreading and sharing, and community-
+based adaptation) and urban areas (e.g., green infrastructure, sustainable land use and planning, and sustainable water 
+management) (medium confidence). {4.3.1, 4.3.2, 4.3.3, 4.3.5, 4.5.3, 4.5.4, 5.3.2, Box 4.2, Box 4.3, Box 4.6, Cross-Chapter 
+Box 9 in Chapter 4}.
+B.6.2 Adaptation is expected to be more challenging for ecosystems, food and health systems at 2°C of global warming than for 
+1.5°C (medium confidence). Some vulnerable regions, including small islands and Least Developed Countries, are projected 
+to experience high multiple interrelated climate risks even at global warming of 1.5°C (high confidence). {3.3.1, 3.4.5,  
+Box 3.5, Table 3.5, Cross-Chapter Box 9 in Chapter 4, 5.6, Cross-Chapter Box 12 in Chapter 5, Box 5.3}
+B.6.3 Limits to adaptive capacity exist at 1.5°C of global warming, become more pronounced at higher levels of warming and 
+vary by sector, with site-specific implications for vulnerable regions, ecosystems and human health (medium confidence).  
+{Cross-Chapter Box 12 in Chapter 5, Box 3.5, Table 3.5} 
+
+SPM Summary for Policymakers1110 Here, impacts on economic growth refer to changes in gross domestic product (GDP). Many impacts, such as loss of human lives, cultural heritage and ecosystem services, are difficult  
+ to value and monetize.1.01.52.0
+01.01.52.00Global mean surface temperature change 
+relative to pre-industrial levels (/zero.numrC)Global mean surface temperature change 
+relative to pre-industrial levels (/zero.numrC)2006-2015How the level of global warming affects impacts and/or risks associated with 
+the Reasons for Concern (RFCs) and selected natural, managed and human 
+systems
+Impacts and risks associated with the Reasons for Concern (RFCs)Purple indicates very high 
+risks of severe impacts/risks 
+and the presence of 
+significant irreversibility or 
+the persistence of 
+climate-related hazards, 
+combined with limited 
+ability to adapt due to the 
+nature of the hazard or 
+impacts/risks. 
+Red indicates severe and 
+widespread impacts/risks. 
+Yellow indicates that 
+impacts/risks are detectable 
+and attributable to climate 
+change with at least medium 
+confidence. 
+White indicates that no 
+impacts are detectable and 
+attributable to climate 
+change.Five Reasons For Concern (RFCs) illustrate the impacts and risks of 
+different levels of global warming for people, economies and ecosystems 
+across sectors and regions.
+Heat-related 
+morbidity 
+and mortalityLevel of additional 
+impact/risk due 
+to climate changeRFC1
+Unique and 
+threatened 
+systemsRFC2
+Extreme 
+weather 
+events RFC4
+Global 
+aggregate 
+impactsRFC5
+Large scale 
+singular 
+eventsRFC3
+Distribution 
+of impacts
+Warm-water
+coralsTerrestrial
+ecosystemsTourism2006-2015
+HVHVHHHH
+HM
+M-HH
+MM
+MM
+M
+HMH
+HH
+MHH
+MM
+HM
+HM
+HM
+HMHImpacts and risks for selected natural, managed and human systems
+Confidence level  for transition: L=Low, M=Medium, H=High and VH=Very highMangroves Small-scale
+low-latitude
+fisheriesArctic
+regionCoastal 
+floodingFluvial 
+floodingCrop 
+yieldsUndetectableModerateHighVery high
+Figure SPM.2 | Five integrative reasons for concern (RFCs) provide a framework for summarizing key impacts and risks across sectors and regions, and were 
+introduced in the IPCC Third Assessment Report. RFCs illustrate the implications of global warming for people, economies and ecosystems. Impacts and/or risks 
+for each RFC are based on assessment of the new literature that has appeared. As in AR5, this literature was used to make expert judgments to assess the levels 
+of global warming at which levels of impact and/or risk are undetectable, moderate, high or very high. The selection of impacts and risks to natural, managed and 
+human systems in the lower panel is illustrative and is not intended to be fully comprehensive. {3.4, 3.5, 3.5.2.1, 3.5.2.2, 3.5.2.3, 3.5.2.4, 3.5.2.5, 5.4.1, 5.5.3, 
+5.6.1, Box 3.4}
+RFC1 Unique and threatened systems: ecological and human systems that have restricted geographic ranges constrained by climate-related conditions and 
+have high endemism or other distinctive properties. Examples include coral reefs, the Arctic and its indigenous people, mountain glaciers and biodiversity hotspots. 
+RFC2 Extreme weather events: risks/impacts to human health, livelihoods, assets and ecosystems from extreme weather events such as heat waves, heavy rain, 
+drought and associated wildfires, and coastal flooding. 
+RFC3 Distribution of impacts: risks/impacts that disproportionately affect particular groups due to uneven distribution of physical climate change hazards, 
+exposure or vulnerability. 
+RFC4 Global aggregate impacts: global monetary damage, global-scale degradation and loss of ecosystems and biodiversity. 
+RFC5 Large-scale singular events: are relatively large, abrupt and sometimes irreversible changes in systems that are caused by global warming. Examples 
+include disintegration of the Greenland and Antarctic ice sheets.
+
+SPMSummary for Policymakers1211 References to pathways limiting global warming to 2°C are based on a 66% probability of staying below 2°C.
+12 Non-CO2 emissions included in this Report are all anthropogenic emissions other than CO2 that result in radiative forcing. These include short-lived climate forcers, such as methane,  
+ some fluorinated gases, ozone precursors, aerosols or aerosol precursors, such as black carbon and sulphur dioxide, respectively, as well as long-lived greenhouse gases, such as nitrous  
+ oxide or some fluorinated gases. The radiative forcing associated with non-CO2 emissions and changes in surface albedo is referred to as non-CO2 radiative forcing. {2.2.1}
+13 There is a clear scientific basis for a total carbon budget consistent with limiting global warming to 1.5°C. However, neither this total carbon budget nor the fraction of this budget  
+ taken up by past emissions were assessed in this Report.
+14 Irrespective of the measure of global temperature used, updated understanding and further advances in methods have led to an increase in the estimated remaining carbon budget of  
+ about 300 GtCO2 compared to AR5. (medium confidence) {2.2.2}
+15 These estimates use observed GMST to 2006–2015 and estimate future temperature changes using near surface air temperatures. C. Emission Pathways and System Transitions Consistent with 1.5°C 
+Global Warming
+C.1  In model pathways with no or limited overshoot of 1.5°C, global net anthropogenic CO2 emissions 
+decline by about 45% from 2010 levels by 2030 (40–60% interquartile range), reaching net zero 
+around 2050 (2045–2055 interquartile range). For limiting global warming to below 2°C11 CO2 
+emissions are projected to decline by about 25% by 2030 in most pathways (10–30% interquartile 
+range) and reach net zero around 2070 (2065–2080 interquartile range). Non-CO2 emissions in 
+pathways that limit global warming to 1.5°C show deep reductions that are similar to those in 
+pathways limiting warming to 2°C. (high confidence) (Figure SPM.3a) {2.1, 2.3, Table 2.4} 
+C.1.1 CO2 emissions reductions that limit global warming to 1.5°C with no or limited overshoot can involve different portfolios of 
+mitigation measures, striking different balances between lowering energy and resource intensity, rate of decarbonization, 
+and the reliance on carbon dioxide removal. Different portfolios face different implementation challenges and potential 
+synergies and trade-offs with sustainable development. (high confidence) (Figure SPM.3b) {2.3.2, 2.3.4, 2.4, 2.5.3}  
+C.1.2 Modelled pathways that limit global warming to 1.5°C with no or limited overshoot involve deep reductions in emissions 
+of methane and black carbon (35% or more of both by 2050 relative to 2010). These pathways also reduce most of the 
+cooling aerosols, which partially offsets mitigation effects for two to three decades. Non-CO2 emissions12 can be reduced 
+as a result of broad mitigation measures in the energy sector. In addition, targeted non-CO2 mitigation measures can 
+reduce nitrous oxide and methane from agriculture, methane from the waste sector, some sources of black carbon, and 
+hydrofluorocarbons. High bioenergy demand can increase emissions of nitrous oxide in some 1.5°C pathways, highlighting 
+the importance of appropriate management approaches. Improved air quality resulting from projected reductions in many 
+non-CO2 emissions provide direct and immediate population health benefits in all 1.5°C model pathways. (high confidence) 
+(Figure SPM.3a) {2.2.1, 2.3.3, 2.4.4, 2.5.3, 4.3.6, 5.4.2} 
+C.1.3 Limiting global warming requires limiting the total cumulative global anthropogenic emissions of CO2 since the pre-
+industrial period, that is, staying within a total carbon budget (high confidence).13 By the end of 2017, anthropogenic CO2 
+emissions since the pre-industrial period are estimated to have reduced the total carbon budget for 1.5°C by approximately 
+2200 ± 320 GtCO2 (medium confidence). The associated remaining budget is being depleted by current emissions of 
+42 ± 3 GtCO2 per year (high confidence). The choice of the measure of global temperature affects the estimated remaining 
+carbon budget. Using global mean surface air temperature, as in AR5, gives an estimate of the remaining carbon budget of 
+580 GtCO2 for a 50% probability of limiting warming to 1.5°C, and 420 GtCO2 for a 66% probability (medium confidence).14 
+Alternatively, using GMST gives estimates of 770 and 570 GtCO2, for 50% and 66% probabilities,15 respectively (medium 
+confidence). Uncertainties in the size of these estimated remaining carbon budgets are substantial and depend on several 
+factors. Uncertainties in the climate response to CO2 and non-CO2 emissions contribute ±400 GtCO2 and the level of historic 
+warming contributes ±250 GtCO2 (medium confidence). Potential additional carbon release from future permafrost thawing 
+and methane release from wetlands would reduce budgets by up to 100 GtCO2 over the course of this century and more 
+thereafter (medium confidence). In addition, the level of non-CO2 mitigation in the future could alter the remaining carbon 
+budget by 250 GtCO2 in either direction (medium confidence). {1.2.4, 2.2.2, 2.6.1, Table 2.2, Chapter 2 Supplementary 
+Material}
+C.1.4 Solar radiation modification (SRM) measures are not included in any of the available assessed pathways. Although some 
+SRM measures may be theoretically effective in reducing an overshoot, they face large uncertainties and knowledge gaps 
+
+SPM Summary for Policymakers13as well as substantial risks and institutional and social constraints to deployment related to governance, ethics, and impacts 
+on sustainable development. They also do not mitigate ocean acidification. (medium confidence) {4.3.8, Cross-Chapter 
+Box 10 in Chapter 4}
+2010 2020 2030 2040 2050 2060 2070 2080 2090 2100
+-20-1001020304050
+Black carbon emissions
+Nitrous oxide emissionsMethane emissionsEmissions of non-CO/two.dnom forcers are also reduced 
+or limited in pathways limiting global warming 
+to 1.5°C with no or limited overshoot, but 
+they do not reach zero globally. Non-CO/two.subs emissions relative to 2010
+Billion tonnes of CO/two.subs/yrGlobal emissions pathway characteristics
+General characteristics of the evolution of anthropogenic net emissions of CO/two.dnom, and total emissions of 
+methane, black carbon, and nitrous oxide in model pathways that limit global warming to 1.5°C with no or 
+limited overshoot. Net emissions are defined as anthropogenic emissions reduced by anthropogenic 
+removals. Reductions in net emissions can be achieved through different portfolios of mitigation measures 
+illustrated in Figure SPM.3b.
+Global total net CO/two.dnom emissions
+2020 2040 2060 2080 2100
+01
+2020 2040 2060 2080 2100
+01
+2020 2040 2060 2080 2100
+01
+Four illustrative model pathways
+In pathways limiting global warming to 1.5°C 
+with no or limited overshoot as well as in 
+pathways with a higher overshoot, CO/two.tnum emissions 
+are reduced to net zero globally around 2050.
+P1
+P2
+P3
+P4
+Pathways with higher overshoot
+Pathways limiting global warming below 2°C
+(Not shown above) Pathways limiting global warming to 1.5°C with no or limited overshoot Timing of net zero CO/two.dnom
+Line widths depict the 5-95th 
+percentile and the 25-75th 
+percentile of scenarios
+Figure SPM.3a | Global emissions pathway characteristics. The main panel shows global net anthropogenic CO2 emissions in pathways limiting global warming 
+to 1.5°C with no or limited (less than 0.1°C) overshoot and pathways with higher overshoot. The shaded area shows the full range for pathways analysed in this 
+Report. The panels on the right show non-CO2 emissions ranges for three compounds with large historical forcing and a substantial portion of emissions coming 
+from sources distinct from those central to CO2 mitigation. Shaded areas in these panels show the 5–95% (light shading) and interquartile (dark shading) ranges 
+of pathways limiting global warming to 1.5°C with no or limited overshoot. Box and whiskers at the bottom of the figure show the timing of pathways reaching 
+global net zero CO2 emission levels, and a comparison with pathways limiting global warming to 2°C with at least 66% probability. Four illustrative model pathways 
+are highlighted in the main panel and are labelled P1, P2, P3 and P4, corresponding to the LED, S1, S2, and S5 pathways assessed in Chapter 2. Descriptions and 
+characteristics of these pathways are available in Figure SPM.3b. {2.1, 2.2, 2.3, Figure 2.5, Figure 2.10, Figure 2.11}
+
+SPMSummary for Policymakers14Breakdown of contributions to global net CO/two.dnom emissions in four illustrative model pathways 
+P1:  A scenario in which social, 
+business and technological innovations 
+result in lower energy demand up to 
+2050 while living standards rise, 
+especially in the global South. A 
+downsized energy system enables 
+rapid decarbonization of energy supply. 
+Afforestation is the only CDR option 
+considered; neither fossil fuels with CCS 
+nor BECCS are used.P2:  A scenario with a broad focus on 
+sustainability including energy 
+intensity, human development, 
+economic convergence and 
+international cooperation, as well as 
+shi/f_ts towards sustainable and healthy 
+consumption patterns, low-carbon 
+technology innovation, and 
+well-managed land systems with 
+limited societal acceptability for BECCS.P3:  A middle-of-the-road scenario in
+which societal as well as technological 
+development follows historical 
+patterns. Emissions reductions are 
+mainly achieved by changing the way in 
+which energy and products are 
+produced, and to a lesser degree by 
+reductions in demand.P4:  A resource- and energy-intensive 
+scenario in which economic growth and 
+globalization lead to widespread 
+adoption of greenhouse-gas-intensive 
+lifestyles, including high demand for 
+transportation fuels and livestock 
+products. Emissions reductions are 
+mainly achieved through technological 
+means, making strong use of CDR 
+through the deployment of BECCS.
+Fossil fuel and industry AFOLU BECCS
+-2002040
+2020 2060 2100-2002040
+2020 2060 2100-2002040
+2020 2060 2100-2002040
+2020 2060 2100
+No or limited overshoot
+-58
+-93
+-50
+-82
+-15
+-32
+60
+77
+-78
+-97
+-37
+-87
+-25
+-74
+59
+150
+-11
+-16
+430
+833
+0
+0
+0.2
+-24
+-33
+5
+6Pathway classification
+CO/two.dnom emission change in 2030 (% rel to 2010)
+               in 2050 (% rel to 2010)
+Kyoto-GHG emissions * in 2030 (% rel to 2010)  
+               in 2050 (% rel to 2010) 
+Final energy demand** in 2030 (% rel to 2010) 
+               in 2050 (% rel to 2010)
+Renewable share in electricity in 2030 (%)
+               in 2050 (%)
+Primary energy from coal in 2030 (% rel to 2010)
+               in 2050 (% rel to 2010)
+     from oil in 2030  (% rel to 2010)
+                in 2050  (% rel to 2010)
+     from gas in 2030  (% rel to 2010)
+                in 2050  (% rel to 2010)
+     from nuclear in 2030  (% rel to 2010)
+                in 2050  (% rel to 2010)
+     from biomass in 2030  (% rel to 2010)
+                in 2050  (% rel to 2010) 
+     from non-biomass renewables in 2030  (% rel to 2010)
+                in 2050  (% rel to 2010)
+Cumulative CCS until 2100 (GtCO/two.dnom)
+               of which BECCS (GtCO/two.dnom)
+Land area of bioenergy crops in 2050 (million km/two.numr)
+Agricultural CH/four.dnom emissions in 2030 (% rel to 2010)
+                in 2050  (% rel to 2010)
+Agricultural N/two.dnomO emissions in 2030 (% rel to 2010)
+                in 2050  (% rel to 2010)
+ No or limited overshoot
+-47
+-95
+-49
+-89
+-5
+2
+58
+81
+-61
+-77
+-13
+-50
+-20
+-53
+83
+98
+0
+49
+470
+1327
+348
+151
+0.9
+-48
+-69
+-26
+-26No or limited overshoot
+-41
+-91
+-35
+-78
+17
+21
+48
+63
+-75
+-73
+-3
+-81
+33
+21
+98
+501
+36
+121
+315
+878
+687
+414
+2.8
+1
+-23
+15
+0Higher overshoot
+4
+-97
+-2
+-80
+39
+44
+25
+70
+-59
+-97
+86
+-32
+37
+-48
+106
+468
+-1
+418
+110
+1137
+1218
+1191
+7.2
+14
+2
+3
+39No or limited overshoot
+(-58,-40)
+(-107,-94)
+(-51,-39)
+(-93,-81)
+(-12,7)
+(-11,22)
+(47,65)
+(69,86)
+(-78, -59) 
+(-95, -74)
+(-34,3)
+(-78,-31)
+(-26,21)
+(-56,6)
+(44,102)
+(91,190)
+(29,80)
+(123,261)
+(245,436)
+(576,1299)
+(550,1017)
+(364,662)
+(1.5,3.2)
+(-30,-11)
+(-47,-24)
+(-21,3)
+(-26,1)Characteristics of four illustrative model pathways
+Different mitigation strategies can achieve the net emissions reductions that would be required to follow a 
+pathway that limits global warming to 1.5°C with no or limited overshoot. All pathways use Carbon Dioxide 
+Removal (CDR), but the amount varies across pathways, as do the relative contributions of Bioenergy with 
+Carbon Capture and Storage (BECCS) and removals in the Agriculture, Forestry and Other Land Use (AFOLU) 
+sector. This has implications for emissions and several other pathway characteristics.
+P1 P2 P3 P4
+P1 P2 P3 P4 Interquartile rangeBillion tonnes CO/two.subs per year (GtCO/two.dnom/yr)
+Global indicatorsBillion tonnes CO/two.subs per year (GtCO/two.dnom/yr) Billion tonnes CO/two.subs per year (GtCO/two.dnom/yr) Billion tonnes CO/two.subs per year (GtCO/two.dnom/yr)
+NOTE: Indicators have been selected to show global trends identified by the Chapter 2 assessment. 
+National and sectoral characteristics can differ substantially from the global trends shown above.* Kyoto-gas emissions are based on IPCC Second Assessment Report GWP-100
+** Changes in energy demand are associated with improvements in energy 
+efficiency and behaviour change
+
+SPM Summary for Policymakers15Figure SPM.3b | Characteristics of four illustrative model pathways in relation to global warming of 1.5°C introduced in Figure SPM.3a. These pathways were 
+selected to show a range of potential mitigation approaches and vary widely in their projected energy and land use, as well as their assumptions about future 
+socio-economic developments, including economic and population growth, equity and sustainability. A breakdown of the global net anthropogenic CO2 emissions 
+into the contributions in terms of CO2 emissions from fossil fuel and industry; agriculture, forestry and other land use (AFOLU); and bioenergy with carbon capture 
+and storage (BECCS) is shown. AFOLU estimates reported here are not necessarily comparable with countries’ estimates. Further characteristics for each of these 
+pathways are listed below each pathway. These pathways illustrate relative global differences in mitigation strategies, but do not represent central estimates, 
+national strategies, and do not indicate requirements. For comparison, the right-most column shows the interquartile ranges across pathways with no or limited 
+overshoot of 1.5°C. Pathways P1, P2, P3 and P4 correspond to the LED, S1, S2 and S5 pathways assessed in Chapter 2 (Figure SPM.3a). {2.2.1, 2.3.1, 2.3.2, 
+2.3.3, 2.3.4, 2.4.1, 2.4.2, 2.4.4, 2.5.3, Figure 2.5, Figure 2.6, Figure 2.9, Figure 2.10, Figure 2.11, Figure 2.14, Figure 2.15, Figure 2.16, Figure 2.17, Figure 2.24,  
+Figure 2.25, Table 2.4, Table 2.6, Table 2.7, Table 2.9, Table 4.1} 
+C.2  Pathways limiting global warming to 1.5°C with no or limited overshoot would require rapid 
+and far-reaching transitions in energy, land, urban and infrastructure (including transport and 
+buildings), and industrial systems (high confidence). These systems transitions are unprecedented 
+in terms of scale, but not necessarily in terms of speed, and imply deep emissions reductions in all 
+sectors, a wide portfolio of mitigation options and a significant upscaling of investments in those 
+options (medium confidence). {2.3, 2.4, 2.5, 4.2, 4.3, 4.4, 4.5}
+C.2.1 Pathways that limit global warming to 1.5°C with no or limited overshoot show system changes that are more rapid and 
+pronounced over the next two decades than in 2°C pathways (high confidence). The rates of system changes associated 
+with limiting global warming to 1.5°C with no or limited overshoot have occurred in the past within specific sectors, 
+technologies and spatial contexts, but there is no documented historic precedent for their scale (medium confidence). 
+{2.3.3, 2.3.4, 2.4, 2.5, 4.2.1, 4.2.2, Cross-Chapter Box 11 in Chapter 4} 
+C.2.2 In energy systems, modelled global pathways (considered in the literature) limiting global warming to 1.5°C with no or 
+limited overshoot (for more details see Figure SPM.3b) generally meet energy service demand with lower energy use, 
+including through enhanced energy efficiency, and show faster electrification of energy end use compared to 2°C (high 
+confidence). In 1.5°C pathways with no or limited overshoot, low-emission energy sources are projected to have a higher 
+share, compared with 2°C pathways, particularly before 2050 (high confidence). In 1.5°C pathways with no or limited 
+overshoot, renewables are projected to supply 70–85% (interquartile range) of electricity in 2050 (high confidence). In 
+electricity generation, shares of nuclear and fossil fuels with carbon dioxide capture and storage (CCS) are modelled to 
+increase in most 1.5°C pathways with no or limited overshoot. In modelled 1.5°C pathways with limited or no overshoot, 
+the use of CCS would allow the electricity generation share of gas to be approximately 8% (3–11% interquartile range) 
+of global electricity in 2050, while the use of coal shows a steep reduction in all pathways and would be reduced to close 
+to 0% (0–2% interquartile range) of electricity (high confidence). While acknowledging the challenges, and differences 
+between the options and national circumstances, political, economic, social and technical feasibility of solar energy, wind 
+energy and electricity storage technologies have substantially improved over the past few years (high confidence). These 
+improvements signal a potential system transition in electricity generation. (Figure SPM.3b) {2.4.1, 2.4.2, Figure 2.1, Table 
+2.6, Table 2.7, Cross-Chapter Box 6 in Chapter 3, 4.2.1, 4.3.1, 4.3.3, 4.5.2}
+C.2.3 CO2 emissions from industry in pathways limiting global warming to 1.5°C with no or limited overshoot are projected to 
+be about 65–90% (interquartile range) lower in 2050 relative to 2010, as compared to 50–80% for global warming of 
+2°C (medium confidence). Such reductions can be achieved through combinations of new and existing technologies and 
+practices, including electrification, hydrogen, sustainable bio-based feedstocks, product substitution, and carbon capture, 
+utilization and storage (CCUS). These options are technically proven at various scales but their large-scale deployment 
+may be limited by economic, financial, human capacity and institutional constraints in specific contexts, and specific 
+characteristics of large-scale industrial installations. In industry, emissions reductions by energy and process efficiency 
+by themselves are insufficient for limiting warming to 1.5°C with no or limited overshoot (high confidence). {2.4.3, 4.2.1, 
+Table 4.1, Table 4.3, 4.3.3, 4.3.4, 4.5.2}
+C.2.4 The urban and infrastructure system transition consistent with limiting global warming to 1.5°C with no or limited overshoot 
+would imply, for example, changes in land and urban planning practices, as well as deeper emissions reductions in transport 
+and buildings compared to pathways that limit global warming below 2°C (medium confidence). Technical measures 
+
+SPMSummary for Policymakers16and practices enabling deep emissions reductions include various energy efficiency options. In pathways limiting global 
+warming to 1.5°C with no or limited overshoot, the electricity share of energy demand in buildings would be about 55–75% 
+in 2050 compared to 50–70% in 2050 for 2°C global warming (medium confidence). In the transport sector, the share of 
+low-emission final energy would rise from less than 5% in 2020 to about 35–65% in 2050 compared to 25–45% for 2°C 
+of global warming (medium confidence). Economic, institutional and socio-cultural barriers may inhibit these urban and 
+infrastructure system transitions, depending on national, regional and local circumstances, capabilities and the availability 
+of capital (high confidence). {2.3.4, 2.4.3, 4.2.1, Table 4.1, 4.3.3, 4.5.2}
+C.2.5 Transitions in global and regional land use are found in all pathways limiting global warming to 1.5°C with no or limited 
+overshoot, but their scale depends on the pursued mitigation portfolio. Model pathways that limit global warming to 1.5°C 
+with no or limited overshoot project a 4 million km2 reduction to a 2.5 million km2 increase of non-pasture agricultural land 
+for food and feed crops and a 0.5–11 million km2 reduction of pasture land, to be converted into a 0–6 million km2 increase 
+of agricultural land for energy crops and a 2 million km2 reduction to 9.5 million km2 increase in forests by 2050 relative 
+to 2010 (medium confidence).16 Land-use transitions of similar magnitude can be observed in modelled 2°C pathways 
+(medium confidence). Such large transitions pose profound challenges for sustainable management of the various demands 
+on land for human settlements, food, livestock feed, fibre, bioenergy, carbon storage, biodiversity and other ecosystem 
+services (high confidence). Mitigation options limiting the demand for land include sustainable intensification of land-use 
+practices, ecosystem restoration and changes towards less resource-intensive diets (high confidence). The implementation 
+of land-based mitigation options would require overcoming socio-economic, institutional, technological, financing and 
+environmental barriers that differ across regions (high confidence). {2.4.4, Figure 2.24, 4.3.2, 4.3.7, 4.5.2, Cross-Chapter 
+Box 7 in Chapter 3}
+C.2.6 Additional annual average energy-related investments for the period 2016 to 2050 in pathways limiting warming to 
+1.5°C compared to pathways without new climate policies beyond those in place today are estimated to be around 830 
+billion USD2010 (range of 150 billion to 1700 billion USD2010 across six models17). This compares to total annual average 
+energy supply investments in 1.5°C pathways of 1460 to 3510 billion USD2010 and total annual average energy demand 
+investments of 640 to 910 billion USD2010 for the period 2016 to 2050. Total energy-related investments increase by 
+about 12% (range of 3% to 24%) in 1.5°C pathways relative to 2°C pathways. Annual investments in low-carbon energy 
+technologies and energy efficiency are upscaled by roughly a factor of six (range of factor of 4 to 10) by 2050 compared to 
+2015 (medium confidence). {2.5.2, Box 4.8, Figure 2.27}
+C.2.7 Modelled pathways limiting global warming to 1.5°C with no or limited overshoot project a wide range of global average 
+discounted marginal abatement costs over the 21st century. They are roughly 3-4 times higher than in pathways limiting 
+global warming to below 2°C (high confidence). The economic literature distinguishes marginal abatement costs from total 
+mitigation costs in the economy. The literature on total mitigation costs of 1.5°C mitigation pathways is limited and was 
+not assessed in this Report. Knowledge gaps remain in the integrated assessment of the economy-wide costs and benefits 
+of mitigation in line with pathways limiting warming to 1.5°C. {2.5.2; 2.6; Figure 2.26}
+16 The projected land-use changes presented are not deployed to their upper limits simultaneously in a single pathway.
+17 Including two pathways limiting warming to 1.5°C with no or limited overshoot and four pathways with higher overshoot.
+
+SPM Summary for Policymakers17C.3  All pathways that limit global warming to 1.5°C with limited or no overshoot project the use of 
+carbon dioxide removal (CDR) on the order of 100–1000 GtCO2 over the 21st century. CDR would 
+be used to compensate for residual emissions and, in most cases, achieve net negative emissions 
+to return global warming to 1.5°C following a peak (high confidence). CDR deployment of several 
+hundreds of GtCO2 is subject to multiple feasibility and sustainability constraints (high confidence). 
+Significant near-term emissions reductions and measures to lower energy and land demand can 
+limit CDR deployment to a few hundred GtCO2 without reliance on bioenergy with carbon capture 
+and storage (BECCS) (high confidence). {2.3, 2.4, 3.6.2, 4.3, 5.4}  
+C.3.1 Existing and potential CDR measures include afforestation and reforestation, land restoration and soil carbon sequestration, 
+BECCS, direct air carbon capture and storage (DACCS), enhanced weathering and ocean alkalinization. These differ widely 
+in terms of maturity, potentials, costs, risks, co-benefits and trade-offs (high confidence). To date, only a few published 
+pathways include CDR measures other than afforestation and BECCS. {2.3.4, 3.6.2, 4.3.2, 4.3.7}
+C.3.2 In pathways limiting global warming to 1.5°C with limited or no overshoot, BECCS deployment is projected to range from 
+0–1, 0–8, and 0–16 GtCO2 yr−1 in 2030, 2050, and 2100, respectively, while agriculture, forestry and land-use (AFOLU) 
+related CDR measures are projected to remove 0–5, 1–11, and 1–5 GtCO2 yr−1 in these years (medium confidence). The 
+upper end of these deployment ranges by mid-century exceeds the BECCS potential of up to 5 GtCO2 yr−1 and afforestation 
+potential of up to 3.6 GtCO2 yr−1 assessed based on recent literature (medium confidence). Some pathways avoid BECCS 
+deployment completely through demand-side measures and greater reliance on AFOLU-related CDR measures (medium 
+confidence). The use of bioenergy can be as high or even higher when BECCS is excluded compared to when it is included 
+due to its potential for replacing fossil fuels across sectors (high confidence). (Figure SPM.3b) {2.3.3, 2.3.4, 2.4.2, 3.6.2, 
+4.3.1, 4.2.3, 4.3.2, 4.3.7, 4.4.3, Table 2.4}
+C.3.3 Pathways that overshoot 1.5°C of global warming rely on CDR exceeding residual CO2 emissions later in the century to 
+return to below 1.5°C by 2100, with larger overshoots requiring greater amounts of CDR (Figure SPM.3b) (high confidence). 
+Limitations on the speed, scale, and societal acceptability of CDR deployment hence determine the ability to return global 
+warming to below 1.5°C following an overshoot. Carbon cycle and climate system understanding is still limited about the 
+effectiveness of net negative emissions to reduce temperatures after they peak (high confidence). {2.2, 2.3.4, 2.3.5, 2.6, 
+4.3.7, 4.5.2, Table 4.11}
+C.3.4 Most current and potential CDR measures could have significant impacts on land, energy, water or nutrients if deployed 
+at large scale (high confidence). Afforestation and bioenergy may compete with other land uses and may have significant 
+impacts on agricultural and food systems, biodiversity, and other ecosystem functions and services (high confidence). 
+Effective governance is needed to limit such trade-offs and ensure permanence of carbon removal in terrestrial, geological 
+and ocean reservoirs (high confidence). Feasibility and sustainability of CDR use could be enhanced by a portfolio of options 
+deployed at substantial, but lesser scales, rather than a single option at very large scale (high confidence). (Figure SPM.3b) 
+{2.3.4, 2.4.4, 2.5.3, 2.6, 3.6.2, 4.3.2, 4.3.7, 4.5.2, 5.4.1, 5.4.2; Cross-Chapter Boxes 7 and 8 in Chapter 3, Table 4.11, Table 
+5.3, Figure 5.3}
+C.3.5 Some AFOLU-related CDR measures such as restoration of natural ecosystems and soil carbon sequestration could provide 
+co-benefits such as improved biodiversity, soil quality, and local food security. If deployed at large scale, they would 
+require governance systems enabling sustainable land management to conserve and protect land carbon stocks and other 
+ecosystem functions and services (medium confidence). (Figure SPM.4) {2.3.3, 2.3.4, 2.4.2, 2.4.4, 3.6.2, 5.4.1, Cross-Chapter 
+Boxes 3 in Chapter 1 and 7 in Chapter 3, 4.3.2, 4.3.7, 4.4.1, 4.5.2, Table 2.4}
+
+SPMSummary for Policymakers18D. Strengthening the Global Response in the Context of Sustainable 
+Development and Efforts to Eradicate Poverty
+D.1 Estimates of the global emissions outcome of current nationally stated mitigation ambitions as 
+submitted under the Paris Agreement would lead to global greenhouse gas emissions18 in 2030 
+of 52–58 GtCO2eq yr−1 (medium confidence). Pathways reflecting these ambitions would not limit 
+global warming to 1.5°C, even if supplemented by very challenging increases in the scale and 
+ambition of emissions reductions after 2030 (high confidence). Avoiding overshoot and reliance 
+on future large-scale deployment of carbon dioxide removal (CDR) can only be achieved if global 
+CO2 emissions start to decline well before 2030 (high confidence). {1.2, 2.3, 3.3, 3.4, 4.2, 4.4, Cross-
+Chapter Box 11 in Chapter 4} 
+D.1.1 Pathways that limit global warming to 1.5°C with no or limited overshoot show clear emission reductions by 2030 (high 
+confidence). All but one show a decline in global greenhouse gas emissions to below 35 GtCO2eq yr−1 in 2030, and half of 
+available pathways fall within the 25–30 GtCO2eq yr−1 range (interquartile range), a 40–50% reduction from 2010 levels 
+(high confidence). Pathways reflecting current nationally stated mitigation ambition until 2030 are broadly consistent 
+with cost-effective pathways that result in a global warming of about 3°C by 2100, with warming continuing afterwards 
+(medium confidence). {2.3.3, 2.3.5, Cross-Chapter Box 11 in Chapter 4, 5.5.3.2}
+D.1.2 Overshoot trajectories result in higher impacts and associated challenges compared to pathways that limit global warming 
+to 1.5°C with no or limited overshoot (high confidence). Reversing warming after an overshoot of 0.2°C or larger during 
+this century would require upscaling and deployment of CDR at rates and volumes that might not be achievable given 
+considerable implementation challenges (medium confidence). {1.3.3, 2.3.4, 2.3.5, 2.5.1, 3.3, 4.3.7, Cross-Chapter Box 8 in 
+Chapter 3, Cross-Chapter Box 11 in Chapter 4}
+D.1.3 The lower the emissions in 2030, the lower the challenge in limiting global warming to 1.5°C after 2030 with no or limited 
+overshoot (high confidence). The challenges from delayed actions to reduce greenhouse gas emissions include the risk of 
+cost escalation, lock-in in carbon-emitting infrastructure, stranded assets, and reduced flexibility in future response options 
+in the medium to long term (high confidence). These may increase uneven distributional impacts between countries at 
+different stages of development (medium confidence). {2.3.5, 4.4.5, 5.4.2}
+D.2 The avoided climate change impacts on sustainable development, eradication of poverty and reducing 
+inequalities would be greater if global warming were limited to 1.5°C rather than 2°C, if mitigation 
+and adaptation synergies are maximized while trade-offs are minimized (high confidence). {1.1, 1.4, 
+2.5, 3.3, 3.4, 5.2, Table 5.1}
+D.2.1 Climate change impacts and responses are closely linked to sustainable development which balances social well-being, 
+economic prosperity and environmental protection. The United Nations Sustainable Development Goals (SDGs), adopted in 
+2015, provide an established framework for assessing the links between global warming of 1.5°C or 2°C and development 
+goals that include poverty eradication, reducing inequalities, and climate action. (high confidence) {Cross-Chapter Box 4 in 
+Chapter 1, 1.4, 5.1}
+D.2.2 The consideration of ethics and equity can help address the uneven distribution of adverse impacts associated with 
+1.5°C and higher levels of global warming, as well as those from mitigation and adaptation, particularly for poor and 
+disadvantaged populations, in all societies (high confidence). {1.1.1, 1.1.2, 1.4.3, 2.5.3, 3.4.10, 5.1, 5.2, 5.3. 5.4, Cross-
+Chapter Box 4 in Chapter 1, Cross-Chapter Boxes 6 and 8 in Chapter 3, and Cross-Chapter Box 12 in Chapter 5}
+D.2.3 Mitigation and adaptation consistent with limiting global warming to 1.5°C are underpinned by enabling conditions, assessed 
+in this Report across the geophysical, environmental-ecological, technological, economic, socio-cultural and institutional 
+18 GHG emissions have been aggregated with 100-year GWP values as introduced in the IPCC Second Assessment Report.
+
+SPM Summary for Policymakers19dimensions of feasibility. Strengthened multilevel governance, institutional capacity, policy instruments, technological 
+innovation and transfer and mobilization of finance, and changes in human behaviour and lifestyles are enabling conditions 
+that enhance the feasibility of mitigation and adaptation options for 1.5°C-consistent systems transitions. (high confidence) 
+{1.4, Cross-Chapter Box 3 in Chapter 1, 2.5.1, 4.4, 4.5, 5.6}
+D.3 Adaptation options specific to national contexts, if carefully selected together with enabling 
+conditions, will have benefits for sustainable development and poverty reduction with global 
+warming of 1.5°C, although trade-offs are possible (high confidence). {1.4, 4.3, 4.5}
+D.3.1 Adaptation options that reduce the vulnerability of human and natural systems have many synergies with sustainable 
+development, if well managed, such as ensuring food and water security, reducing disaster risks, improving health 
+conditions, maintaining ecosystem services and reducing poverty and inequality (high confidence). Increasing investment 
+in physical and social infrastructure is a key enabling condition to enhance the resilience and the adaptive capacities 
+of societies. These benefits can occur in most regions with adaptation to 1.5°C of global warming (high confidence).  
+{1.4.3, 4.2.2, 4.3.1, 4.3.2, 4.3.3, 4.3.5, 4.4.1, 4.4.3, 4.5.3, 5.3.1, 5.3.2}
+D.3.2 Adaptation to 1.5°C global warming can also result in trade-offs or maladaptations with adverse impacts for sustainable 
+development. For example, if poorly designed or implemented, adaptation projects in a range of sectors can increase 
+greenhouse gas emissions and water use, increase gender and social inequality, undermine health conditions, and encroach 
+on natural ecosystems (high confidence). These trade-offs can be reduced by adaptations that include attention to poverty 
+and sustainable development (high confidence). {4.3.2, 4.3.3, 4.5.4, 5.3.2; Cross-Chapter Boxes 6 and 7 in Chapter 3} 
+D.3.3 A mix of adaptation and mitigation options to limit global warming to 1.5°C, implemented in a participatory and integrated 
+manner, can enable rapid, systemic transitions in urban and rural areas (high confidence). These are most effective when 
+aligned with economic and sustainable development, and when local and regional governments and decision makers are 
+supported by national governments (medium confidence). {4.3.2, 4.3.3, 4.4.1, 4.4.2}
+D.3.4 Adaptation options that also mitigate emissions can provide synergies and cost savings in most sectors and system 
+transitions, such as when land management reduces emissions and disaster risk, or when low-carbon buildings are also 
+designed for efficient cooling. Trade-offs between mitigation and adaptation, when limiting global warming to 1.5°C, 
+such as when bioenergy crops, reforestation or afforestation encroach on land needed for agricultural adaptation, can 
+undermine food security, livelihoods, ecosystem functions and services and other aspects of sustainable development. (high 
+confidence) {3.4.3, 4.3.2, 4.3.4, 4.4.1, 4.5.2, 4.5.3, 4.5.4}
+D.4 Mitigation options consistent with 1.5°C pathways are associated with multiple synergies and trade-
+offs across the Sustainable Development Goals (SDGs). While the total number of possible synergies 
+exceeds the number of trade-offs, their net effect will depend on the pace and magnitude of changes, 
+the composition of the mitigation portfolio and the management of the transition. (high confidence) 
+(Figure SPM.4) {2.5, 4.5, 5.4} 
+D.4.1 1.5°C pathways have robust synergies particularly for the SDGs 3 (health), 7 (clean energy), 11 (cities and communities), 12 
+(responsible consumption and production) and 14 (oceans) (very high confidence). Some 1.5°C pathways show potential 
+trade-offs with mitigation for SDGs 1 (poverty), 2 (hunger), 6 (water) and 7 (energy access), if not managed carefully (high 
+confidence). (Figure SPM.4) {5.4.2; Figure 5.4, Cross-Chapter Boxes 7 and 8 in Chapter 3}  
+D.4.2 1.5°C pathways that include low energy demand (e.g., see P1 in Figure SPM.3a and SPM.3b), low material consumption, 
+and low GHG-intensive food consumption have the most pronounced synergies and the lowest number of trade-offs with 
+respect to sustainable development and the SDGs (high confidence). Such pathways would reduce dependence on CDR. In 
+modelled pathways, sustainable development, eradicating poverty and reducing inequality can support limiting warming to 
+1.5°C (high confidence). (Figure SPM.3b, Figure SPM.4) {2.4.3, 2.5.1, 2.5.3, Figure 2.4, Figure 2.28, 5.4.1, 5.4.2, Figure 5.4} 
+
+SPMSummary for Policymakers20Indicative linkages between mitigation options and sustainable 
+development using SDGs (The linkages do not show costs and benefits)
+Mitigation options deployed in each sector can be associated with potential positive effects (synergies) or 
+negative effects (trade-offs) with the Sustainable Development Goals (SDGs). The degree to which this 
+potential is realized will depend on the selected portfolio of mitigation options, mitigation policy design, 
+and local circumstances and context. Particularly in the energy-demand sector, the potential for synergies is 
+larger than for trade-offs. The bars group individually assessed options by level of confidence and take into 
+account the relative strength of the assessed mitigation-SDG connections.
+The overall size of the coloured bars depict the relative 
+potential for synergies and trade-offs between the sectoral 
+mitigation options and the SDGs.Length shows strength of connection
+Energy Supply Land
+Trade-offs               Synergies Trade-offs              Synergies Trade-offs              Synergies
+The shades depict the level of confidence of the 
+assessed potential for Trade-offs/Synergies.
+Very High LowShades show level of confidence
+Energy Demand 
+SDG1
+No Poverty
+SDG2
+Zero Hunger
+SDG 3
+Good Health
+and Well-being
+SDG 4
+Quality
+Education
+SDG 5
+Gender
+Equality
+SDG 6
+Clean Water
+and Sanitation
+SDG 7
+Affordable and
+Clean Energy
+SDG 8
+Decent Work
+and Economic
+Growth
+SDG 9
+Industry,
+Innovation and
+Infrastructure
+SDG 10
+Reduced
+Inequalities
+SDG 11
+Sustainable
+Cities and
+Communities
+SDG 12
+Responsible
+Consumption
+and Production
+SDG 14
+Life Below
+Water
+SDG 15
+Life on Land
+SDG 16
+Peace, Justice
+and Strong
+Institutions
+SDG 17
+Partnerships for
+ the Goals
+
+SPM Summary for Policymakers21D.4.3 1.5°C and 2°C modelled pathways often rely on the deployment of large-scale land-related measures like afforestation 
+and bioenergy supply, which, if poorly managed, can compete with food production and hence raise food security concerns 
+(high confidence). The impacts of carbon dioxide removal (CDR) options on SDGs depend on the type of options and the 
+scale of deployment (high confidence). If poorly implemented, CDR options such as BECCS and AFOLU options would lead 
+to trade-offs. Context-relevant design and implementation requires considering people’s needs, biodiversity, and other 
+sustainable development dimensions (very high confidence). (Figure SPM.4) {5.4.1.3, Cross-Chapter Box 7 in Chapter 3} 
+D.4.4 Mitigation consistent with 1.5°C pathways creates risks for sustainable development in regions with high dependency on 
+fossil fuels for revenue and employment generation (high confidence). Policies that promote diversification of the economy 
+and the energy sector can address the associated challenges (high confidence). {5.4.1.2, Box 5.2} 
+D.4.5 Redistributive policies across sectors and populations that shield the poor and vulnerable can resolve trade-offs for a range 
+of SDGs, particularly hunger, poverty and energy access. Investment needs for such complementary policies are only a small 
+fraction of the overall mitigation investments in 1.5°C pathways. (high confidence) {2.4.3, 5.4.2, Figure 5.5} 
+D.5 Limiting the risks from global warming of 1.5°C in the context of sustainable development and 
+poverty eradication implies system transitions that can be enabled by an increase of adaptation 
+and mitigation investments, policy instruments, the acceleration of technological innovation and 
+behaviour changes (high confidence). {2.3, 2.4, 2.5, 3.2, 4.2, 4.4, 4.5, 5.2, 5.5, 5.6}
+D.5.1 Directing finance towards investment in infrastructure for mitigation and adaptation could provide additional resources.  
+This could involve the mobilization of private funds by institutional investors, asset managers and development or 
+investment banks, as well as the provision of public funds. Government policies that lower the risk of low-emission and 
+adaptation investments can facilitate the mobilization of private funds and enhance the effectiveness of other public 
+policies. Studies indicate a number of challenges, including access to finance and mobilization of funds. (high confidence) 
+{2.5.1, 2.5.2, 4.4.5} 
+D.5.2 Adaptation finance consistent with global warming of 1.5°C is difficult to quantify and compare with 2°C. Knowledge 
+gaps include insufficient data to calculate specific climate resilience-enhancing investments from the provision of currently 
+underinvested basic infrastructure. Estimates of the costs of adaptation might be lower at global warming of 1.5°C than for 
+2°C. Adaptation needs have typically been supported by public sector sources such as national and subnational government 
+budgets, and in developing countries together with support from development assistance, multilateral development banks, 
+and United Nations Framework Convention on Climate Change channels (medium confidence). More recently there is a Figure SPM.4 | Potential synergies and trade-offs between the sectoral portfolio of climate change mitigation options and the Sustainable Development Goals 
+(SDGs). The SDGs serve as an analytical framework for the assessment of the different sustainable development dimensions, which extend beyond the time frame 
+of the 2030 SDG targets. The assessment is based on literature on mitigation options that are considered relevant for 1.5°C. The assessed strength of the SDG 
+interactions is based on the qualitative and quantitative assessment of individual mitigation options listed in Table 5.2. For each mitigation option, the strength of 
+the SDG-connection as well as the associated confidence of the underlying literature (shades of green and red) was assessed. The strength of positive connections 
+(synergies) and negative connections (trade-offs) across all individual options within a sector (see Table 5.2) are aggregated into sectoral potentials for the whole 
+mitigation portfolio. The (white) areas outside the bars, which indicate no interactions, have low confidence due to the uncertainty and limited number of studies 
+exploring indirect effects. The strength of the connection considers only the effect of mitigation and does not include benefits of avoided impacts. SDG 13 (climate 
+action) is not listed because mitigation is being considered in terms of interactions with SDGs and not vice versa. The bars denote the strength of the connection, 
+and do not consider the strength of the impact on the SDGs. The energy demand sector comprises behavioural responses, fuel switching and efficiency options in 
+the transport, industry and building sector as well as carbon capture options in the industry sector. Options assessed in the energy supply sector comprise biomass 
+and non-biomass renewables, nuclear, carbon capture and storage (CCS) with bioenergy, and CCS with fossil fuels. Options in the land sector comprise agricultural 
+and forest options, sustainable diets and reduced food waste, soil sequestration, livestock and manure management, reduced deforestation, afforestation and 
+reforestation, and responsible sourcing. In addition to this figure, options in the ocean sector are discussed in the underlying report. {5.4, Table 5.2, Figure 5.2}
+Information about the net impacts of mitigation on sustainable development in 1.5°C pathways is available only for a limited number of SDGs and mitigation 
+options. Only a limited number of studies have assessed the benefits of avoided climate change impacts of 1.5°C pathways for the SDGs, and the co-effects 
+of adaptation for mitigation and the SDGs. The assessment of the indicative mitigation potentials in Figure SPM.4 is a step further from AR5 towards a more 
+comprehensive and integrated assessment in the future.
+
+SPMSummary for Policymakers22growing understanding of the scale and increase in non-governmental organizations and private funding in some regions 
+(medium confidence). Barriers include the scale of adaptation financing, limited capacity and access to adaptation finance 
+(medium confidence). {4.4.5, 4.6} 
+D.5.3 Global model pathways limiting global warming to 1.5°C are projected to involve the annual average investment needs 
+in the energy system of around 2.4 trillion USD2010 between 2016 and 2035, representing about 2.5% of the world GDP 
+(medium confidence). {4.4.5, Box 4.8}
+D.5.4 Policy tools can help mobilize incremental resources, including through shifting global investments and savings and 
+through market and non-market based instruments as well as accompanying measures to secure the equity of the 
+transition, acknowledging the challenges related with implementation, including those of energy costs, depreciation of 
+assets and impacts on international competition, and utilizing the opportunities to maximize co-benefits (high confidence).  
+{1.3.3, 2.3.4, 2.3.5, 2.5.1, 2.5.2, Cross-Chapter Box 8 in Chapter 3, Cross-Chapter Box 11 in Chapter 4, 4.4.5, 5.5.2}
+D.5.5 The systems transitions consistent with adapting to and limiting global warming to 1.5°C include the widespread adoption 
+of new and possibly disruptive technologies and practices and enhanced climate-driven innovation. These imply enhanced 
+technological innovation capabilities, including in industry and finance. Both national innovation policies and international 
+cooperation can contribute to the development, commercialization and widespread adoption of mitigation and adaptation 
+technologies. Innovation policies may be more effective when they combine public support for research and development 
+with policy mixes that provide incentives for technology diffusion. (high confidence) {4.4.4, 4.4.5}.  
+D.5.6 Education, information, and community approaches, including those that are informed by indigenous knowledge and local 
+knowledge, can accelerate the wide-scale behaviour changes consistent with adapting to and limiting global warming to 
+1.5°C. These approaches are more effective when combined with other policies and tailored to the motivations, capabilities 
+and resources of specific actors and contexts (high confidence). Public acceptability can enable or inhibit the implementation 
+of policies and measures to limit global warming to 1.5°C and to adapt to the consequences. Public acceptability depends 
+on the individual’s evaluation of expected policy consequences, the perceived fairness of the distribution of these 
+consequences, and perceived fairness of decision procedures (high confidence). {1.1, 1.5, 4.3.5, 4.4.1, 4.4.3, Box 4.3, 5.5.3, 
+5.6.5} 
+D.6 Sustainable development supports, and often enables, the fundamental societal and systems 
+transitions and transformations that help limit global warming to 1.5°C. Such changes facilitate the 
+pursuit of climate-resilient development pathways that achieve ambitious mitigation and adaptation 
+in conjunction with poverty eradication and efforts to reduce inequalities (high confidence). {Box 1.1, 
+1.4.3, Figure 5.1, 5.5.3, Box 5.3} 
+D.6.1 Social justice and equity are core aspects of climate-resilient development pathways that aim to limit global warming to 
+1.5°C as they address challenges and inevitable trade-offs, widen opportunities, and ensure that options, visions, and values 
+are deliberated, between and within countries and communities, without making the poor and disadvantaged worse off 
+(high confidence). {5.5.2, 5.5.3, Box 5.3, Figure 5.1, Figure 5.6, Cross-Chapter Boxes 12 and 13 in Chapter 5}
+D.6.2 The potential for climate-resilient development pathways differs between and within regions and nations, due to different 
+development contexts and systemic vulnerabilities (very high confidence). Efforts along such pathways to date have been 
+limited (medium confidence) and enhanced efforts would involve strengthened and timely action from all countries and 
+non-state actors (high confidence). {5.5.1, 5.5.3, Figure 5.1}
+D.6.3 Pathways that are consistent with sustainable development show fewer mitigation and adaptation challenges and are 
+associated with lower mitigation costs. The large majority of modelling studies could not construct pathways characterized 
+by lack of international cooperation, inequality and poverty that were able to limit global warming to 1.5°C. (high 
+confidence) {2.3.1, 2.5.1, 2.5.3, 5.5.2}
+
+SPM Summary for Policymakers23D.7 Strengthening the capacities for climate action of national and sub-national authorities, civil society, 
+the private sector, indigenous peoples and local communities can support the implementation of 
+ambitious actions implied by limiting global warming to 1.5°C (high confidence). International 
+cooperation can provide an enabling environment for this to be achieved in all countries and for all 
+people, in the context of sustainable development. International cooperation is a critical enabler for 
+developing countries and vulnerable regions (high confidence). {1.4, 2.3, 2.5, 4.2, 4.4, 4.5, 5.3, 5.4, 5.5, 
+5.6, 5, Box 4.1, Box 4.2, Box 4.7, Box 5.3, Cross-Chapter Box 9 in Chapter 4, Cross-Chapter Box 13 in 
+Chapter 5}
+D.7.1 Partnerships involving non-state public and private actors, institutional investors, the banking system, civil society and 
+scientific institutions would facilitate actions and responses consistent with limiting global warming to 1.5°C (very high 
+confidence). {1.4, 4.4.1, 4.2.2, 4.4.3, 4.4.5, 4.5.3, 5.4.1, 5.6.2, Box 5.3}.
+D.7.2 Cooperation on strengthened accountable multilevel governance that includes non-state actors such as industry, civil 
+society and scientific institutions, coordinated sectoral and cross-sectoral policies at various governance levels, gender-
+sensitive policies, finance including innovative financing, and cooperation on technology development and transfer can 
+ensure participation, transparency, capacity building and learning among different players (high confidence). {2.5.1, 2.5.2, 
+4.2.2, 4.4.1, 4.4.2, 4.4.3, 4.4.4, 4.4.5, 4.5.3, Cross-Chapter Box 9 in Chapter 4, 5.3.1, 5.5.3, Cross-Chapter Box 13 in Chapter 
+5, 5.6.1, 5.6.3}
+D.7.3 International cooperation is a critical enabler for developing countries and vulnerable regions to strengthen their action for 
+the implementation of 1.5°C-consistent climate responses, including through enhancing access to finance and technology 
+and enhancing domestic capacities, taking into account national and local circumstances and needs (high confidence). 
+{2.3.1, 2.5.1, 4.4.1, 4.4.2, 4.4.4, 4.4.5, 5.4.1 5.5.3, 5.6.1, Box 4.1, Box 4.2, Box 4.7}.
+D.7.4 Collective efforts at all levels, in ways that reflect different circumstances and capabilities, in the pursuit of limiting global 
+warming to 1.5°C, taking into account equity as well as effectiveness, can facilitate strengthening the global response to 
+climate change, achieving sustainable development and eradicating poverty (high confidence). {1.4.2, 2.3.1, 2.5.1, 2.5.2, 
+2.5.3, 4.2.2, 4.4.1, 4.4.2, 4.4.3, 4.4.4, 4.4.5, 4.5.3, 5.3.1, 5.4.1, 5.5.3, 5.6.1, 5.6.2, 5.6.3}
+
+SPMSummary for Policymakers24Box SPM.1:  Core Concepts Central to this Special Report 
+Global mean surface temperature (GMST): Estimated global average of near-surface air temperatures over land and 
+sea ice, and sea surface temperatures over ice-free ocean regions, with changes normally expressed as departures from a 
+value over a specified reference period. When estimating changes in GMST, near-surface air temperature over both land 
+and oceans are also used.19 {1.2.1.1} 
+Pre-industrial: The multi-century period prior to the onset of large-scale industrial activity around 1750. The reference 
+period 1850–1900 is used to approximate pre-industrial GMST. {1.2.1.2} 
+Global warming: The estimated increase in GMST averaged over a 30-year period, or the 30-year period centred on a 
+particular year or decade, expressed relative to pre-industrial levels unless otherwise specified. For 30-year periods that 
+span past and future years, the current multi-decadal warming trend is assumed to continue. {1.2.1}
+Net zero CO2 emissions: Net zero carbon dioxide (CO2) emissions are achieved when anthropogenic CO2 emissions are 
+balanced globally by anthropogenic CO2 removals over a specified period. 
+Carbon dioxide removal (CDR): Anthropogenic activities removing CO2 from the atmosphere and durably storing it in 
+geological, terrestrial, or ocean reservoirs, or in products. It includes existing and potential anthropogenic enhancement of 
+biological or geochemical sinks and direct air capture and storage, but excludes natural CO2 uptake not directly caused by 
+human activities.
+Total carbon budget: Estimated cumulative net global anthropogenic CO2 emissions from the pre-industrial period 
+to the time that anthropogenic CO2 emissions reach net zero that would result, at some probability, in limiting global 
+warming to a given level, accounting for the impact of other anthropogenic emissions. {2.2.2} 
+Remaining carbon budget: Estimated cumulative net global anthropogenic CO2 emissions from a given start date to the 
+time that anthropogenic CO2 emissions reach net zero that would result, at some probability, in limiting global warming 
+to a given level, accounting for the impact of other anthropogenic emissions. {2.2.2}
+Temperature overshoot: The temporary exceedance of a specified level of global warming. 
+Emission pathways: In this Summary for Policymakers, the modelled trajectories of global anthropogenic emissions over 
+the 21st century are termed emission pathways. Emission pathways are classified by their temperature trajectory over 
+the 21st century: pathways giving at least 50% probability based on current knowledge of limiting global warming to 
+below 1.5°C are classified as ‘no overshoot’; those limiting warming to below 1.6°C and returning to 1.5°C by 2100 are 
+classified as ‘1.5°C limited-overshoot’; while those exceeding 1.6°C but still returning to 1.5°C by 2100 are classified as 
+‘higher-overshoot’.
+Impacts: Effects of climate change on human and natural systems. Impacts can have beneficial or adverse outcomes 
+for livelihoods, health and well-being, ecosystems and species, services, infrastructure, and economic, social and cultural 
+assets.
+Risk: The potential for adverse consequences from a climate-related hazard for human and natural systems, resulting 
+from the interactions between the hazard and the vulnerability and exposure of the affected system. Risk integrates 
+the likelihood of exposure to a hazard and the magnitude of its impact. Risk also can describe the potential for adverse 
+consequences of adaptation or mitigation responses to climate change. 
+Climate-resilient development pathways (CRDPs): Trajectories that strengthen sustainable development at multiple 
+scales and efforts to eradicate poverty through equitable societal and systems transitions and transformations while 
+reducing the threat of climate change through ambitious mitigation, adaptation and climate resilience. 
+19 Past IPCC reports, reflecting the literature, have used a variety of approximately equivalent metrics of GMST change.
+

WG1AR5_all_final.txt

@@ -0,0 +1,899 @@
+CLIMATE CHANGE 2013
+The Physical Science Basis
+WORKING GROUP I CONTRIBUTION TO THE 
+FIFTH ASSESSMENT REPORT OF THE 
+INTERGOVERNMENTAL PANEL ON CLIMATE CHANGE
+WG IINTERGOVERNMENTAL PANEL ON  climate change
+
+ 
+
+ForewordClimate Change 2013
+The Physical Science Basis
+Working Group I Contribution to the 
+Fifth Assessment Report of the
+Intergovernmental Panel on Climate Change
+Edited by
+ Thomas F. Stocker Dahe Qin 
+ Working Group I Co-Chair Working Group I Co-Chair
+ University of Bern China Meteorological Administration
+ Gian-Kasper Plattner  Melinda M.B. Tignor Simon K. Allen Judith Boschung
+ Director of Science Director of Operations Senior Science Officer Administrative Assistant 
+  
+ Alexander Nauels Yu Xia Vincent Bex Pauline M. Midgley
+ Science Assistant Science Officer IT Officer Head   
+  
+Working Group I Technical Support Unit
+
+ii
+ForewordCAMBRIDGE UNIVERSITY PRESS
+Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, São Paolo, Delhi, Mexico City
+Cambridge University Press
+32 Avenue of the Americas, New York, NY 10013-2473, USA
+www.cambridge.org
+Information on this title: www.cambridge.org/9781107661820
+© Intergovernmental Panel on Climate Change 2013
+This publication is in copyright. Subject to statutory exception and to the provisions of relevant collective licensing agreements,
+no reproduction of any part may take place without the written permission of Cambridge University Press.
+First published 2013
+Printed in the United States of America
+A catalog record for this publication is available from the British Library.
+ISBN 978-1-107-05799-1 hardback
+ISBN 978-1-107-66182-0 paperback
+Cambridge University Press has no responsibility for the persistence or accuracy of URLs for external or third-party Internet Web
+sites referred to in this publication and does not guarantee that any content on such Web sites is, or will remain, accurate or appropriate.
+Please use the following reference to the whole report:
+IPCC, 2013: Climate Change 2013: The Physical Science Basis. Contribution of Working Group I to the Fifth Assessment Report of the Intergovern -
+mental Panel on Climate Change [Stocker, T.F ., D. Qin, G.-K. Plattner, M. Tignor, S.K. Allen, J. Boschung, A. Nauels, Y . Xia, V. Bex and P .M. Midgley 
+(eds.)]. Cambridge University Press, Cambridge, United Kingdom and New York, NY , USA, 1535 pp.
+Cover photo:
+Folgefonna glacier on the high plateaus of Sørfjorden, Norway (60°03’ N - 6°20’ E) © Yann Arthus-Bertrand / Altitude.
+
+Introduction Chapter 2iii
+ForewordForeword, Preface 
+and Dedication
+
+v
+ForewordForeword
+“Climate Change 2013: The Physical Science Basis” presents clear and 
+robust conclusions in a global assessment of climate change science—
+not the least of which is that the science now shows with 95 percent 
+certainty that human activity is the dominant cause of observed warm -
+ing since the mid-20th century. The report confirms that warming in 
+the climate system is unequivocal, with many of the observed changes 
+unprecedented over decades to millennia: warming of the atmosphere 
+and the ocean, diminishing snow and ice, rising sea levels and increas -
+ing concentrations of greenhouse gases. Each of the last three decades 
+has been successively warmer at the Earth’s surface than any preced -
+ing decade since 1850. 
+These and other findings confirm and enhance our scientific under -
+standing of the climate system and the role of greenhouse gas emis -
+sions; as such, the report demands the urgent attention of both policy -
+makers and the general public.
+As an intergovernmental body jointly established in 1988 by the World 
+Meteorological Organization (WMO) and the United Nations Environ -
+ment Programme (UNEP), the Intergovernmental Panel on Climate 
+Change (IPCC) has provided policymakers with the most authorita -
+tive and objective scientific and technical assessments. Beginning in 
+1990, this series of IPCC Assessment Reports, Special Reports, Tech -
+nical Papers, Methodology Reports and other products have become 
+standard works of reference. 
+This Working Group I contribution to the IPCC’s Fifth Assessment 
+Report contains important new scientific knowledge that can be used 
+to produce climate information and services for assisting society to act 
+to address the challenges of climate change. The timing is particularly 
+significant, as this information provides a new impetus, through clear 
+and indisputable physical science, to those negotiators responsible for 
+concluding a new agreement under the United Nations Framework 
+Convention on Climate Change in 2015.
+Climate change is a long-term challenge, but one that requires urgent 
+action given the pace and the scale by which greenhouse gases are 
+accumulating in the atmosphere and the risks of a more than 2 degree 
+Celsius temperature rise. Today we need to focus on the fundamentals 
+and on the actions otherwise the risks we run will get higher with 
+every year. 
+This Working Group I assessment was made possible thanks to the 
+commitment and dedication of many hundreds of experts worldwide, 
+representing a wide range of disciplines. WMO and UNEP are proud 
+that so many of the experts belong to their communities and networks. 
+We express our deep gratitude to all authors, review editors and expert 
+reviewers for devoting their knowledge, expertise and time. We would 
+like to thank the staff of the Working Group I Technical Support Unit 
+and the IPCC Secretariat for their dedication. We are also grateful to the governments that supported their scien -
+tists’ participation in developing this report and that contributed to 
+the IPCC Trust Fund to provide for the essential participation of experts 
+from developing countries and countries with economies in transition. 
+We would like to express our appreciation to the government of Italy 
+for hosting the scoping meeting for the IPCC’s Fifth Assessment Report, 
+to the governments of China, France, Morocco and Australia for host -
+ing drafting sessions of the Working Group I contribution and to the 
+government of Sweden for hosting the Twelfth Session of Working 
+Group I in Stockholm for approval of the Working Group I Report. The 
+generous financial support by the government of Switzerland, and the 
+logistical support by the University of Bern (Switzerland), enabled the 
+smooth operation of the Working Group I Technical Support Unit. This 
+is gratefully acknowledged.
+We would particularly like to thank Dr. Rajendra Pachauri, Chairman of 
+the IPCC, for his direction and guidance of the IPCC and we express our 
+deep gratitude to Professor Qin Dahe and Professor Thomas Stocker, 
+the Co-Chairs of Working Group I for their tireless leadership through -
+out the development and production of this report. 
+M. Jarraud
+Secretary-General
+World Meteorological Organization
+A. Steiner
+Executive Director
+United Nations Environment Programme
+
+
+vii
+PrefacePreface
+The Working Group I contribution to the Fifth Assessment Report of 
+the Intergovernmental Panel on Climate Change (IPCC) provides a 
+comprehensive assessment of the physical science basis of climate 
+change. It builds upon the Working Group I contribution to the IPCC’s 
+Fourth Assessment Report in 2007 and incorporates subsequent new 
+findings from the Special Report on Managing the Risks of Extreme 
+Events and Disasters to Advance Climate Change Adaptation, as well 
+as from research published in the extensive scientific and technical 
+literature. The assessment considers new evidence of past, present and 
+projected future climate change based on many independent scien -
+tific analyses from observations of the climate system, paleoclimate 
+archives, theoretical studies of climate processes and simulations using 
+climate models.
+ 
+Scope of the Report
+During the process of scoping and approving the outline of its Fifth 
+Assessment Report, the IPCC focussed on those aspects of the current 
+understanding of the science of climate change that were judged to be 
+most relevant to policymakers.
+In this report, Working Group I has extended coverage of future climate 
+change compared to earlier reports by assessing near-term projections 
+and predictability as well as long-term projections and irreversibility 
+in two separate chapters. Following the decisions made by the Panel 
+during the scoping and outline approval, a set of new scenarios, the 
+Representative Concentration Pathways, are used across all three 
+Working Groups for projections of climate change over the 21st cen -
+tury. The coverage of regional information in the Working Group I 
+report is expanded by specifically assessing climate phenomena such 
+as monsoon systems and their relevance to future climate change in 
+the regions. 
+The Working Group I Report is an assessment, not a review or a text 
+book of climate science, and is based on the published scientific and 
+technical literature available up to 15 March 2013. Underlying all 
+aspects of the report is a strong commitment to assessing the science 
+comprehensively, without bias and in a way that is relevant to policy 
+but not policy prescriptive. 
+Structure of the Report
+This report consists of a short Summary for Policymakers, a longer 
+Technical Summary and fourteen thematic chapters plus annexes. An 
+innovation in this Working Group I assessment is the Atlas of Global 
+and Regional Climate Projections (Annex I) containing time series and 
+maps of temperature and precipitation projections for 35 regions of 
+the world, which enhances accessibility for stakeholders and users.The Summary for Policymakers and Technical Summary of this report 
+follow a parallel structure and each includes cross-references to the 
+chapter and section where the material being summarised can be 
+found in the underlying report. In this way, these summary compo -
+nents of the report provide a road-map to the contents of the entire 
+report and a traceable account of every major finding.
+In order to facilitate the accessibility of the findings of the Working 
+Group I assessment for a wide readership and to enhance their usabil -
+ity for stakeholders, each section of the Summary for Policymakers has 
+a highlighted headline statement. Taken together, these 19 headline 
+statements provide an overarching summary in simple and quotable 
+language that is supported by the scientists and approved by the 
+member governments of the IPCC. Another innovative feature of this 
+report is the presentation of Thematic Focus Elements in the Techni -
+cal Summary that provide end to end assessments of important cross-
+cutting issues in the physical science basis of climate change. 
+Introduction (Chapter 1):  This chapter provides information on the 
+progress in climate change science since the First Assessment Report 
+of the IPCC in 1990 and gives an overview of key concepts, indica -
+tors of climate change, the treatment of uncertainties and advances in 
+measurement and modelling capabilities. This includes a description of 
+the future scenarios and in particular the Representative Concentration 
+Pathway scenarios used across all Working Groups for the IPCC’s Fifth 
+Assessment Report. 
+Observations and Paleoclimate Information (Chapters 2, 3, 4, 5):  These 
+chapters assess information from all climate system components on 
+climate variability and change as obtained from instrumental records 
+and climate archives. They cover all relevant aspects of the atmosphere 
+including the stratosphere, the land surface, the oceans and the cryo -
+sphere. Timescales from days to decades (Chapters 2, 3 and 4) and 
+from centuries to many millennia (Chapter 5) are considered.
+Process Understanding (Chapters 6 and 7):  These chapters cover all 
+relevant aspects from observations and process understanding to pro -
+jections from global to regional scales for two key topics. Chapter 6 
+covers the carbon cycle and its interactions with other biogeochemical 
+cycles, in particular the nitrogen cycle, as well as feedbacks on the 
+climate system. For the first time, there is a chapter dedicated to the 
+assessment of the physical science basis of clouds and aerosols, their 
+interactions and chemistry, and the role of water vapour, as well as 
+their role in feedbacks on the climate system (Chapter 7).
+From Forcing to Attribution of Climate Change (Chapters 8, 9, 10): All 
+the information on the different drivers (natural and anthropogenic) 
+of climate change is collected, expressed in terms of Radiative Forc -
+ing and assessed in Chapter 8. In Chapter 9, the hierarchy of climate 
+models used in simulating past and present climate change is assessed 
+and evaluated against observations and paleoclimate reconstructions. 
+
+Preface  viii
+PrefaceInformation regarding detection of changes on global to regional 
+scales and their attribution to the increase in anthropogenic green -
+house gases is assessed in Chapter 10.
+Future Climate Change, Predictability and Irreversibility (Chapters 11 
+and 12):  These chapters assess projections of future climate change 
+derived from climate models on time scales from decades to centuries 
+at both global and regional scales, including mean changes, variabil -
+ity and extremes. Fundamental questions related to the predictability 
+of climate as well as long term climate change, climate change com -
+mitments and inertia in the climate system are addressed. Knowledge 
+on irreversible changes and surprises in the climate system is also 
+assessed.
+Integration (Chapters 13 and 14):  These chapters synthesise all relevant 
+information for two key topics of this assessment: sea level change 
+(Chapter 13) and climate phenomena across the regions (Chapter 14). 
+Chapter 13 presents an end to end assessment of information on sea 
+level change based on paleoclimate reconstructions, observations and 
+process understanding, and provides projections from global to region -
+al scales. Chapter 14 assesses the most important modes of variability 
+in the climate system, such as El Niño-Southern Oscillation, monsoon 
+and many others, as well as extreme events. Furthermore, this chapter 
+deals with interconnections between the climate phenomena, their 
+regional expressions and their relevance for future regional climate 
+change. 
+Maps assessed in Chapter 14, together with Chapters 11 and 12, form 
+the basis of the Atlas of Global and Regional Climate Projections in 
+Annex I, which is also available in digital format. Radiative forcings 
+and estimates of future atmospheric concentrations from Chapters 7, 
+8, 11 and 12 form the basis of the Climate System Scenario Tables 
+presented in Annex II. All material including high-resolution versions of 
+the figures, underlying data and Supplementary Material to the chap -
+ters is also available online: www.climatechange2013.org.
+The scientific community and the climate modelling centres around the 
+world brought together their activities in the Coordinated Modelling 
+Intercomparison Project Phase 5 (CMIP5), providing the basis for most 
+of the assessment of future climate change in this report. Their efforts 
+enable Working Group I to deliver comprehensive scientific informa -
+tion for the policymakers and the users of this report, as well as for 
+the specific assessments of impacts carried out by IPCC Working Group 
+II, and of costs and mitigation strategies, carried out by IPCC Working 
+Group III.
+Following the successful introduction in the previous Working Group I 
+assessment in 2007, all chapters contain Frequently Asked Questions. 
+In these the authors provide scientific answers to a range of general 
+questions in a form that will be accessible to a broad readership and 
+serves as a resource for teaching purposes. Finally, the report is accom -
+panied by extensive Supplementary Material which is made available in the online versions of the report to provide an additional level of 
+detail, such as description of datasets, models, or methodologies used 
+in chapter analyses, as well as material supporting the figures in the 
+Summary for Policymakers.
+The Process
+This Working Group I Assessment Report represents the combined 
+efforts of hundreds of leading experts in the field of climate science 
+and has been prepared in accordance with rules and procedures estab -
+lished by the IPCC. A scoping meeting for the Fifth Assessment Report 
+was held in July 2009 and the outlines for the contributions of the 
+three Working Groups were approved at the 31st Session of the Panel 
+in November 2009. Governments and IPCC observer organisations 
+nominated experts for the author team. The team of 209 Coordinat -
+ing Lead Authors and Lead Authors plus 50 Review Editors selected 
+by the Working Group I Bureau was accepted at the 41st Session of 
+the IPCC Bureau in May 2010. In addition, more than 600 Contribut -
+ing Authors provided draft text and information to the author teams 
+at their request. Drafts prepared by the authors were subject to two 
+rounds of formal review and revision followed by a final round of gov -
+ernment comments on the Summary for Policymakers. A total of 54,677 
+written review comments were submitted by 1089 individual expert 
+reviewers and 38 governments. The Review Editors for each chapter 
+monitored the review process to ensure that all substantive review 
+comments received appropriate consideration. The Summary for Poli -
+cymakers was approved line-by-line and the underlying chapters were 
+then accepted at the 12th Session of IPCC Working Group I from 23–27 
+September 2007.
+Acknowledgements
+We are very grateful for the expertise, hard work, commitment to 
+excellence and integrity shown throughout by the Coordinating Lead 
+Authors and Lead Authors with important help by the many Contribut -
+ing Authors. The Review Editors have played a critical role in assist -
+ing the author teams and ensuring the integrity of the review process. 
+We express our sincere appreciation to all the expert and government 
+reviewers. We would also like to thank the members of the Bureau of 
+Working Group I: Jean Jouzel, Abdalah Mokssit, Fatemeh Rahimizadeh, 
+Fredolin Tangang, David Wratt and Francis Zwiers, for their thoughtful 
+advice and support throughout the preparation of the report. 
+We gratefully acknowledge the long-term efforts of the scientific com -
+munity, organized and facilitated through the World Climate Research 
+Programme, in particular CMIP5. In this effort by climate modelling 
+centres around the world, more than 2 million gigabytes of numerical 
+data have been produced, which were archived and distributed under 
+the stewardship of the Program for Climate Model Diagnosis and Inter -
+comparison. This represents an unprecedented concerted effort by the 
+scientific community and their funding institutions.
+
+ Prefaceix
+PrefaceOur sincere thanks go to the hosts and organizers of the four Working 
+Group I Lead Author Meetings and the 12th Session of Working Group 
+I. We gratefully acknowledge the support from the host countries: 
+China, France, Morocco, Australia and Sweden. The support for their 
+scientists provided by many governments as well as through the IPCC 
+Trust Fund is much appreciated. The efficient operation of the Working 
+Group I Technical Support Unit was made possible by the generous 
+financial support provided by the government of Switzerland and logis -
+tical support from the University of Bern (Switzerland). 
+We would also like to thank Renate Christ, Secretary of the IPCC, and 
+the staff of the IPCC Secretariat: Gaetano Leone, Jonathan Lynn, Mary 
+Jean Burer, Sophie Schlingemann, Judith Ewa, Jesbin Baidya, Werani 
+Zabula, Joelle Fernandez, Annie Courtin, Laura Biagioni and Amy 
+Smith. Thanks are due to Francis Hayes who served as the conference 
+officer for the Working Group I Approval Session.
+Rajendra K. Pachauri Qin Dahe Thomas F. Stocker
+IPCC Chair IPCC WGI Co-Chair IPCC WGI Co-Chair
+Finally our particular appreciation goes to the Working Group I Techni -
+cal Support Unit: Gian-Kasper Plattner, Melinda Tignor, Simon Allen, 
+Judith Boschung, Alexander Nauels, Yu Xia, Vincent Bex and Pauline 
+Midgley for their professionalism, creativity and dedication. Their tire -
+less efforts to coordinate the Working Group I Report ensured a final 
+product of high quality. They were assisted in this by Adrien Michel 
+and Flavio Lehner with further support from Zhou Botao and Sun Ying. 
+In addition, the following contributions are gratefully acknowledged: 
+David Hansford (editorial assistance with the Frequently Asked Ques -
+tions), UNEP/GRID-Geneva and University of Geneva (graphics assis -
+tance with the Frequently Asked Questions), Theresa Kornak (copyedit), 
+Marilyn Anderson (index) and Michael Shibao (design and layout).
+
+xi
+DedicationDedication
+Bert Bolin  
+(15 May 1925 – 30 December 2007)
+The Working Group I contribution to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC) 
+Climate Change 2013: The Physical Science Basis  is dedicated to the memory of Bert Bolin, the first Chair of the IPCC. 
+As an accomplished scientist who published on both atmospheric dynamics and the carbon cycle, including processes in the 
+atmosphere, oceans and biosphere, Bert Bolin realised the complexity of the climate system and its sensitivity to anthropogenic 
+perturbation. He made a fundamental contribution to the organisation of international cooperation in climate research, being 
+involved in the establishment of a number of global programmes.
+Bert Bolin played a key role in the creation of the IPCC and its assessments, which are carried out in a unique and formalized 
+process in order to provide a robust scientific basis for informed decisions regarding one of the greatest challenges of our time. 
+His vision and leadership of the Panel as the founding Chair from 1988 to 1997 laid the basis for subsequent assessments includ -
+ing this one and are remembered with deep appreciation.
+
+
+ForewordContents
+ Front Matter  Foreword    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
+ Preface    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  vii
+ Dedication    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  xi
+ SPM  Summary for Policymakers     . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   3
+ TS  Technical Summary     . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   33
+ Chapters  Chapter 1   Introduction    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   119
+ Chapter 2   Observations : Atmosphere and Surface     . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   159
+ Chapter 3   Observations: Ocean    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   255
+ Chapter 4   Observations: Cryosphere     . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   317
+ Chapter 5   Information from Paleoclimate Archives    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   383
+ Chapter 6   Carbon and Other Biogeochemical Cycles    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   465
+ Chapter 7   Clouds and Aerosols    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   571
+ Chapter 8   Anthropogenic and Natural Radiative Forcing    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   659
+ Chapter 9   Evaluation of Climate Models    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   741
+ Chapter 10   Detection and Attribution of Climate Change: from Global to Regional    . . . . . . . . . . . . . . . .   867
+ Chapter 11   Near-term Climate Change: Projections and Predictability    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   953
+ Chapter 12   Long-term Climate Change: Projections, Commitments and Irreversibility    . . . . . . . . . . . .   1029
+ Chapter 13   Sea Level Change    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   1137
+ Chapter 14   Climate Phenomena and their Relevance for Future Regional Climate Change    . . . . . .   1217
+ Annexes  Annex I   Atlas of Global and Regional Climate Projections    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  1311
+ Annex II   Climate System Scenario Tables    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  1395
+ Annex III   Glossary   . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  1447
+ Annex IV   Acronyms   . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  1467
+ Annex V   Contributors to the IPCC WGI Fifth Assessment Report   . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  1477
+ Annex VI   Expert Reviewers of the IPCC WGI Fifth Assessment Report   . . . . . . . . . . . . . . . . . . . . . . . . . . .  1497
+ Index      . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  1523
+
+Introduction Chapter 2
+Chapter 1Summary for Policymakers
+
+3
+1This Summary for Policymakers should be cited as:
+IPCC, 2013: Summary for Policymakers. In: Climate Change 2013: The Physical Science Basis.  Contribution of 
+Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change  [Stocker, 
+T.F ., D. Qin, G.-K. Plattner, M. Tignor, S.K. Allen, J. Boschung, A. Nauels, Y . Xia, V. Bex and P .M. Midgley (eds.)]. 
+Cambridge University Press, Cambridge, United Kingdom and New York, NY , USA.Summary  
+for Policymakers SPM
+Drafting Authors:
+Lisa V. Alexander (Australia), Simon K. Allen (Switzerland/New Zealand), Nathaniel L. Bindoff 
+(Australia), François-Marie Bréon (France), John A. Church (Australia), Ulrich Cubasch 
+(Germany), Seita Emori (Japan), Piers Forster (UK), Pierre Friedlingstein (UK/Belgium), Nathan 
+Gillett (Canada), Jonathan M. Gregory (UK), Dennis L. Hartmann (USA), Eystein Jansen 
+(Norway), Ben Kirtman (USA), Reto Knutti (Switzerland), Krishna Kumar Kanikicharla (India), 
+Peter Lemke (Germany), Jochem Marotzke (Germany), Valérie Masson-Delmotte (France), 
+Gerald A. Meehl (USA), Igor I. Mokhov (Russian Federation), Shilong Piao (China), Gian-Kasper 
+Plattner (Switzerland), Qin Dahe (China), Venkatachalam Ramaswamy (USA), David Randall 
+(USA), Monika Rhein (Germany), Maisa Rojas (Chile), Christopher Sabine (USA), Drew Shindell 
+(USA), Thomas F . Stocker (Switzerland), Lynne  D. Talley (USA), David G. Vaughan (UK), Shang-
+Ping Xie (USA)
+Draft Contributing Authors:
+Myles R. Allen (UK), Olivier Boucher (France), Don Chambers (USA), Jens Hesselbjerg Christensen 
+(Denmark), Philippe Ciais (France), Peter U. Clark (USA), Matthew Collins (UK), Josefino C. 
+Comiso (USA), Viviane Vasconcellos de Menezes (Australia/Brazil), Richard A. Feely (USA), 
+Thierry Fichefet (Belgium), Arlene M. Fiore (USA), Gregory Flato (Canada), Jan Fuglestvedt 
+(Norway), Gabriele Hegerl (UK/Germany), Paul J. Hezel (Belgium/USA), Gregory C. Johnson 
+(USA), Georg Kaser (Austria/Italy), Vladimir Kattsov (Russian Federation), John Kennedy (UK), 
+Albert M. G. Klein Tank (Netherlands), Corinne Le Quéré (UK), Gunnar Myhre (Norway), Timothy 
+Osborn (UK), Antony J. Payne (UK), Judith Perlwitz (USA), Scott Power (Australia), Michael 
+Prather (USA), Stephen R. Rintoul (Australia), Joeri Rogelj (Switzerland/Belgium), Matilde 
+Rusticucci (Argentina), Michael Schulz (Germany), Jan Sedláček (Switzerland), Peter A. Stott 
+(UK), Rowan Sutton (UK), Peter W. Thorne (USA/Norway/UK), Donald Wuebbles (USA)
+
+SPMSummary for Policymakers41 In this Summary for Policymakers, the following summary terms are used to describe the available evidence: limited, medium, or robust; and for the degree of agreement: 
+low, medium, or high. A level of confidence is expressed using five qualifiers: very low, low, medium, high, and very high, and typeset in italics, e.g., medium confidence . 
+For a given evidence and agreement statement, different confidence levels can be assigned, but increasing levels of evidence and degrees of agreement are correlated with 
+increasing confidence (see Chapter 1 and Box TS.1 for more details).
+2 In this Summary for Policymakers, the following terms have been used to indicate the assessed likelihood of an outcome or a result: virtually certain 99–100% probability, 
+very likely 90–100%, likely 66–100%, about as likely as not 33–66%, unlikely 0–33%, very unlikely 0–10%, exceptionally unlikely 0–1%. Additional terms (extremely likely: 
+95–100%, more likely than not >50–100%, and extremely unlikely 0–5%) may also be used when appropriate. Assessed likelihood is typeset in italics, e.g., very likely  (see 
+Chapter 1 and Box TS.1 for more details).Warming of the climate system is unequivocal, and since the 1950s, many of the observed 
+changes are unprecedented over decades to millennia. The atmosphere and ocean have 
+warmed, the amounts of snow and ice have diminished, sea level has risen, and the 
+concentrations of greenhouse gases have increased (see Figures SPM.1, SPM.2, SPM.3 and 
+SPM.4). {2.2, 2.4, 3.2, 3.7, 4.2–4.7, 5.2, 5.3, 5.5–5.6, 6.2, 13.2}A. Introduction
+The Working Group I contribution to the IPCC’s Fifth Assessment Report (AR5) considers new evidence of climate change 
+based on many independent scientific analyses from observations of the climate system, paleoclimate archives, theoretical 
+studies of climate processes and simulations using climate models. It builds upon the Working Group I contribution to the 
+IPCC’s Fourth Assessment Report (AR4), and incorporates subsequent new findings of research. As a component of the 
+fifth assessment cycle, the IPCC Special Report on Managing the Risks of Extreme Events and Disasters to Advance Climate 
+Change Adaptation (SREX) is an important basis for information on changing weather and climate extremes.
+This Summary for Policymakers (SPM) follows the structure of the Working Group I report. The narrative is supported by a 
+series of overarching highlighted conclusions which, taken together, provide a concise summary. Main sections are introduced 
+with a brief paragraph in italics which outlines the methodological basis of the assessment.
+The degree of certainty in key findings in this assessment is based on the author teams’ evaluations of underlying scientific 
+understanding and is expressed as a qualitative level of confidence (from very low  to very high ) and, when possible, 
+probabilistically with a quantified likelihood (from exceptionally unlikely  to virtually certain ). Confidence in the validity of 
+a finding is based on the type, amount, quality, and consistency of evidence (e.g., data, mechanistic understanding, theory, 
+models, expert judgment) and the degree of agreement1. Probabilistic estimates of quantified measures of uncertainty in a 
+finding are based on statistical analysis of observations or model results, or both, and expert judgment2. Where appropriate, 
+findings are also formulated as statements of fact without using uncertainty qualifiers. (See Chapter 1 and Box TS.1 for more 
+details about the specific language the IPCC uses to communicate uncertainty).
+The basis for substantive paragraphs in this Summary for Policymakers can be found in the chapter sections of the underlying 
+report and in the Technical Summary. These references are given in curly brackets. 
+B. Observed Changes in the Climate System
+Observations of the climate system are based on direct measurements and remote sensing from satellites and other platforms. 
+Global-scale observations from the instrumental era began in the mid-19th century for temperature and other variables, with 
+more comprehensive and diverse sets of observations available for the period 1950 onwards. Paleoclimate reconstructions 
+extend some records back hundreds to millions of years. Together, they provide a comprehensive view of the variability and 
+long-term changes in the atmosphere, the ocean, the cryosphere, and the land surface.
+
+SPM Summary for Policymakers5Each of the last three decades has been successively warmer at the Earth’s surface than any 
+preceding decade since 1850 (see Figure SPM.1). In the Northern Hemisphere, 1983–2012 
+was likely  the warmest 30-year period of the last 1400 years ( medium confidence ). {2.4, 5.3}B.1 Atmosphere
+• The globally averaged combined land and ocean surface temperature data as calculated by a linear trend, show a 
+warming of 0.85 [0.65 to 1.06] °C3, over the period 1880 to 2012, when multiple independently produced datasets exist.  
+The total increase between the average of the 1850–1900 period and the 2003–2012 period is 0.78 [0.72 to 0.85] °C, 
+based on the single longest dataset available4 (see Figure SPM.1). {2.4}
+• For the longest period when calculation of regional trends is sufficiently complete (1901 to 2012), almost the entire globe 
+has experienced surface warming (see Figure SPM.1). {2.4}
+• In addition to robust multi-decadal warming, global mean surface temperature exhibits substantial decadal and 
+interannual variability (see Figure SPM.1). Due to natural variability, trends based on short records are very sensitive to 
+the beginning and end dates and do not in general reflect long-term climate trends. As one example, the rate of warming 
+over the past 15 years (1998–2012; 0.05 [–0.05 to 0.15] °C per decade), which begins with a strong El Niño, is smaller 
+than the rate calculated since 1951 (1951–2012; 0.12 [0.08 to 0.14] °C per decade)5. {2.4}
+• Continental-scale surface temperature reconstructions show, with high confidence , multi-decadal periods during 
+the Medieval Climate Anomaly (year 950 to 1250) that were in some regions as warm as in the late 20th century. 
+These regional warm periods did not occur as coherently across regions as the warming in the late 20th century  (high 
+confidence ). {5.5}
+• It is virtually certain that globally the troposphere has warmed since the mid-20th century. More complete observations 
+allow greater confidence in estimates of tropospheric temperature changes in the extratropical Northern Hemisphere 
+than elsewhere. There is medium confidence  in the rate of warming and its vertical structure in the Northern Hemisphere 
+extra-tropical troposphere and low confidence  elsewhere. {2.4}
+• Confidence  in precipitation change averaged over global land areas since 1901 is low prior to 1951 and medium  
+afterwards.  Averaged over the mid-latitude land areas of the Northern Hemisphere, precipitation has increased since 
+1901 ( medium confidence  before and high confidence  after 1951). For other latitudes area-averaged long-term positive 
+or negative trends have low confidence  (see Figure SPM.2). {TS TFE.1, Figure 2; 2.5}
+• Changes in many extreme weather and climate events have been observed since   about 1950 (see Table SPM.1 for 
+details). It is very likely  that the number of cold days and nights has decreased and the number of warm days and nights 
+has increased on the global scale6. It is likely  that the frequency of heat waves has increased in large parts of Europe, 
+Asia and Australia. There are likely  more land regions where the number of heavy precipitation events has increased than 
+where it has decreased. The frequency or intensity of heavy precipitation events has likely  increased in North America and 
+Europe. In other continents, confidence  in changes in heavy precipitation events is at most medium . {2.6}
+3 In the WGI contribution to the AR5, uncertainty is quantified using 90% uncertainty intervals unless otherwise stated. The 90% uncertainty interval, reported in square 
+brackets, is expected to have a 90% likelihood of covering the value that is being estimated. Uncertainty intervals are not necessarily symmetric about the corresponding 
+best estimate. A best estimate of that value is also given where available. 
+4 Both methods presented in this bullet were also used in AR4. The first calculates the difference using a best fit linear trend of all points between 1880 and 2012. The second 
+calculates the difference between averages for the two periods 1850–1900 and 2003–2012. Therefore, the resulting values and their 90% uncertainty intervals are not 
+directly comparable. {2.4}
+5 Trends for 15-year periods starting in 1995, 1996, and 1997 are 0.13 [0.02 to 0.24] °C per decade, 0.14 [0.03 to 0.24] °C per decade, and, 0.07 [–0.02 to 0.18] °C per 
+decade, respectively.
+6 See the Glossary for the definition of these terms: cold days/cold nights, warm days/warm nights, heat waves.
+
+SPMSummary for Policymakers6Figure SPM.1 |  (a) Observed global mean combined land and ocean surface temperature anomalies, from 1850 to 2012 from three data sets. Top panel: 
+annual mean values. Bottom panel: decadal mean values including the estimate of uncertainty for one dataset (black). Anomalies are relative to the mean 
+of 1961−1990. (b) Map of the observed surface temperature change from 1901 to 2012 derived from temperature trends determined by linear regression 
+from one dataset (orange line in panel a). Trends have been calculated where data availability permits a robust estimate (i.e., only for grid boxes with 
+greater than 70% complete records and more than 20% data availability in the first and last 10% of the time period). Other areas are white. Grid boxes 
+where the trend is significant at the 10% level are indicated by a + sign. For a listing of the datasets and further technical details see the Technical Summary 
+Supplementary Material. {Figures 2.19–2.21; Figure TS.2}
+Temperature anomaly (°C) relative to 1961–1990(a)
+(b) Observed change in surface temperature 1901–2012    −0.6−0.4−0.20.00.20.40.6
+Annual average
+    −0.6−0.4−0.20.00.20.40.6
+1850 1900 1950 2000Decadal average
+(°C) Observed globally averaged combined land and ocean 
+surface temperature anomaly 1850–2012
+−0.6 −0.4 −0.2 00 .2 0.40 .6 0.81 .0 1.25 1.51 .752 .5Year
+
+SPM Summary for Policymakers7Phenomenon and 
+direction of trendAssessment that changes occurred (typically 
+since 1950 unless otherwise indicated)Assessment of a human  
+contribution to observed changes Early 21st century Late 21st century
+Warmer and/or fewer 
+cold days and nights 
+over most land areasVery likely  {2.6}
+Very likely
+Very likely  Very likely  {10.6}
+Likely
+Likely   Likely  {11.3} Virtually certain  {12.4}
+Virtually certain
+Virtually certain  
+Warmer and/or more 
+frequent hot days and  
+nights over most land areasVery likely {2.6}
+Very likely
+Very likelyVery likely  {10.6}
+Likely
+Likely (nights only)Likely  {11.3} Virtually certain  {12.4}
+Virtually certain
+Virtually certain
+Warm spells/heat waves.  
+Frequency and/or duration  
+increases over most 
+land areasMedium confidence  on a global scale  
+Likely in large parts of Europe, Asia and Australia {2.6}
+Medium confidence in many (but not all) regions
+LikelyLikelya
+ {10.6}
+Not formally assessed
+More likely than notNot formally assessedb
+ {11.3}Very likely 
+ {12.4}
+Very likely
+Very likely
+Heavy precipitation events.
+Increase in the frequency, 
+intensity, and/or amount 
+of heavy precipitationLikely more land areas with increases than decreasesc 
+ {2.6}
+Likely more  land areas with increases than decreases
+Likely  over most land areasMedium confidence  
+ {7.6, 10.6}
+Medium confidence
+More likely than notLikely over many land areas  
+ {11.3}Very likely  over most of the mid-latitude land  
+masses and over wet tropical regions  {12.4}
+Likely over many areas
+Very likely over most land areas
+Increases in intensity  
+and/or duration of droughtLow confidence  on a global scale  
+Likely changes in some regionsd {2.6}
+Medium confidence in some regions
+Likely  in many regions, since 1970e  Low confidence  {10.6}
+Medium confidencef
+More likely than notLow confidenceg {11.3} Likely  (medium confidence) on a regional to  
+global scaleh  {12.4}
+Medium confidence in some regions
+Likelye
+Increases in intense 
+tropical cyclone activityLow confidence  in long term (centennial) changes  
+Virtually certain in North Atlantic since 1970 {2.6}
+Low confidence
+Likely  in some regions, since 1970  Low confidencei  
+ {10.6}
+Low confidence
+More likely than notLow confidence   
+ {11.3}More likely than not  in the Western North Pacific  
+and North Atlanticj {14.6}
+More likely than not in some basins
+Likely
+Increased incidence and/or  
+magnitude of extreme 
+high sea level Likely (since 1970) {3.7}
+Likely (late 20th century)
+Likely  Likelyk {3.7}
+Likelyk
+More likely than notkLikelyl {13.7} Very likelyl {13.7}
+Very likelym
+LikelyLikelihood of further changesTable SPM.1 |  Extreme weather and climate events: Global-scale assessment of recent observed changes, human contribution to the changes, and projected further changes for the early (2016–2035) and late (2081–2100) 21st century. 
+Bold indicates where the AR5 (black) provides a revised* global-scale assessment from the SREX (blue) or AR4 (red). Projections for early 21st century were not provided in previous assessment reports. Projections in the AR5 are relative to 
+the reference period of 1986–2005, and use the new Representative Concentration Pathway (RCP) scenarios (see Box SPM.1) unless otherwise specified. See the Glossary for definitions of extreme weather and climate events.
+* The direct comparison of assessment findings between reports is difficult. For some climate variables, different aspects have been assessed, and the revised guidance note on uncertainties has been used for the SREX and AR5. The availability of new information, improved scientific understanding, continued 
+analyses of data and models, and specific differences in methodologies applied in the assessed studies, all contribute to revised assessment findings.
+Notes:
+a Attribution is based on available case studies. It is likely  that human influence has more than doubled the probability of occurrence of some observed heat waves in some locations.
+b Models project near-term increases in the duration, intensity and spatial extent of heat waves and warm spells.
+c In most continents, confidence  in trends is not higher than medium  except in North America and Europe where there have been likely  increases in either the frequency or intensity of heavy precipitation with some seasonal and/or regional variation. It is very likely  that there have been increases in central 
+North America.
+d The frequency and intensity of drought has  likely  increased in the Mediterranean and West Africa, and likely  decreased in central North America and north-west Australia.
+e AR4 assessed the area affected by drought. 
+f SREX assessed medium confidence  that anthropogenic influence had contributed to some changes in the drought patterns observed in the second half of the 20th century, based on its attributed impact on precipitation and temperature changes. SREX assessed low confidence  in the attribution of changes 
+in droughts at the level of single regions.
+g There is  low confidence  in projected changes in soil moisture.
+h Regional to global-scale projected decreases in soil moisture and increased agricultural drought are likely (medium confidence)  in presently dry regions by the end of this century under the RCP8.5 scenario. Soil moisture drying in the Mediterranean, Southwest US and southern African regions is consistent 
+with projected changes in Hadley circulation and increased surface temperatures, so there is high confidence  in likely  surface drying in these regions by the end of this century under the RCP8.5 scenario.
+i There is medium confidence  that a reduction in aerosol forcing over the North Atlantic has contributed at least in part to the observed increase in tropical cyclone activity since the 1970s in this region.
+j Based on expert judgment and assessment of projections which use an SRES A1B (or similar) scenario.
+k Attribution is based on the close relationship between observed changes in extreme and mean sea level.
+l There is high confidence  that this increase in extreme high sea level will primarily be the result of an increase in mean sea level. There is low confidence  in region-specific projections of storminess and associated storm surges.
+m SREX assessed it to be very likely  that mean sea level rise will contribute to future upward trends in extreme coastal high water levels.
+
+SPMSummary for Policymakers8B.2 Ocean
+Ocean warming dominates the increase in energy stored in the climate system, accounting 
+for more than 90% of the energy accumulated between 1971 and 2010 ( high confidence ). 
+It is virtually certain  that the upper ocean (0−700 m) warmed from 1971 to 2010 (see Figure 
+SPM.3), and it likely  warmed between the 1870s and 1971. {3.2, Box 3.1}
+• On a global scale, the ocean warming is largest near the surface, and the upper 75 m warmed by 0.11 [0.09 to 0.13] °C 
+per decade over the period 1971 to 2010. Since AR4, instrumental biases in upper-ocean temperature records have been 
+identified and reduced, enhancing  confidence in the assessment of change. {3.2}
+• It is likely  that the ocean warmed between 700 and 2000 m from 1957 to 2009. Sufficient observations are available for 
+the period 1992 to 2005 for a global assessment of temperature change below 2000 m. There were likely  no significant 
+observed temperature trends between 2000 and 3000 m for this period. It is likely that the ocean warmed from 3000 m 
+to the bottom for this period, with the largest warming observed in the Southern Ocean. {3.2}
+• More than 60% of the net energy increase in the climate system is stored in the upper ocean (0–700 m) during the 
+relatively well-sampled 40-year period from 1971 to 2010, and about 30% is stored in the ocean below 700 m. The 
+increase in upper ocean heat content during this time period estimated from a linear trend is likely  17 [15 to 19] ×  
+1022 J 7 (see Figure SPM.3). {3.2, Box 3.1} 
+• It is about as  likely as not  that ocean heat content from 0–700 m increased more slowly during 2003 to 2010 than during 
+1993 to 2002 (see Figure SPM.3). Ocean heat uptake from 700–2000 m, where interannual variability is smaller, likely  
+continued unabated from 1993 to 2009. {3.2, Box 9.2}
+• It is very likely that regions of high salinity where evaporation dominates have become more saline, while regions of 
+low salinity where precipitation dominates have become fresher since the 1950s. These regional trends in ocean salinity 
+provide indirect evidence that evaporation and precipitation over the oceans have changed ( medium confidence ). {2.5, 
+3.3, 3.5}
+• There is no observational evidence of a trend in the Atlantic Meridional Overturning Circulation (AMOC), based on the 
+decade-long record of the complete AMOC and longer records of individual AMOC components. {3.6} Figure SPM.2 |  Maps of observed precipitation change from 1901 to 2010 and from 1951 to 2010 (trends in annual accumulation calculated using the 
+same criteria as in Figure SPM.1) from one data set. For further technical details see the Technical Summary Supplementary Material. {TS TFE.1, Figure 2; 
+Figure 2.29} −100 −50 −25 −10 −5 −2.5 0 2.5 51 02 55 0 100
+(mm yr-1 per decade)1901– 2010 1951– 2010Observed change in annual precipitation over land
+7 A constant supply of heat through the ocean surface at the rate of 1 W m–2 for 1 year would increase the ocean heat content by 1.1 × 1022 J.
+
+SPM Summary for Policymakers9B.3 Cryosphere
+Over the last two decades, the Greenland and Antarctic ice sheets have been losing mass, 
+glaciers have continued to shrink almost worldwide, and Arctic sea ice and Northern 
+Hemisphere spring snow cover have continued to decrease in extent ( high confidence ) (see 
+Figure SPM.3). {4.2–4.7}
+• The average rate of ice loss8 from glaciers around the world, excluding glaciers on the periphery of the ice sheets9, was 
+very likely  226 [91 to 361] Gt yr−1 over the period 1971 to 2009, and very likely  275 [140 to 410] Gt yr−1 over the period 
+1993 to 200910. {4.3}
+• The average rate of ice loss from the Greenland ice sheet has very likely  substantially increased from 34 [–6 to 74] Gt yr–1 
+over the period 1992 to 2001 to 215 [157 to 274] Gt yr–1 over the period 2002 to 2011. {4.4}
+• The average rate of ice loss from the Antarctic ice sheet has likely increased from 30 [–37 to 97] Gt yr–1 over the period 
+1992–2001 to 147 [72 to 221] Gt yr–1 over the period 2002 to 2011. There is very high confidence  that these losses are 
+mainly from the northern Antarctic Peninsula and the Amundsen Sea sector of West Antarctica. {4.4}
+• The annual mean Arctic sea ice extent decreased over the period 1979 to 2012 with a rate that was very likely  in the 
+range 3.5 to 4.1% per decade (range of 0.45 to 0.51 million km2 per decade), and  very likely  in the range 9.4 to 13.6% 
+per decade (range of 0.73 to 1.07 million km2 per decade) for the summer sea ice minimum (perennial sea ice). The 
+average decrease in decadal mean extent of Arctic sea ice has been most rapid in summer ( high confidence ); the spatial 
+extent has decreased in every season, and in every  successive decade since 1979 ( high confidence ) (see Figure SPM.3). 
+There is medium confidence  from reconstructions that over the past three decades, Arctic summer sea ice retreat was 
+unprecedented and sea surface temperatures were anomalously high in at least the last 1,450 years. {4.2, 5.5}
+• It is very likely  that the annual mean Antarctic sea ice extent increased at a rate in the range of 1.2 to 1.8% per decade 
+(range of 0.13 to 0.20 million km2 per decade) between 1979 and 2012. There is high confidence  that there are strong 
+regional differences in this annual rate, with extent increasing in some regions and decreasing in others. {4.2}
+• There is very high confidence  that the extent of Northern Hemisphere snow cover has decreased since the mid-20th 
+century (see Figure SPM.3). Northern Hemisphere snow cover extent decreased 1.6 [0.8 to 2.4] % per decade for March 
+and April, and 11.7 [8.8 to 14.6] % per decade for June, over the 1967 to 2012 period. During this period, snow cover 
+extent in the Northern Hemisphere did not show a statistically significant increase in any month. {4.5}
+• There is high confidence  that permafrost temperatures have increased in most regions since the early 1980s. Observed 
+warming was up to 3°C in parts of Northern Alaska (early 1980s to mid-2000s) and up to 2°C in parts of the Russian 
+European North (1971 to 2010). In the latter region, a considerable reduction in permafrost thickness and areal extent 
+has been observed over the period 1975 to 2005 ( medium confidence ). {4.7}
+• Multiple lines of evidence support very substantial Arctic warming since the mid-20th century. {Box 5.1, 10.3}
+8 All references to ‘ice loss’ or ‘mass loss’ refer to net ice loss, i.e., accumulation minus melt and iceberg calving. 
+9 For methodological reasons, this assessment of ice loss from the Antarctic and Greenland ice sheets includes change in the glaciers on the periphery. These peripheral glaciers 
+are thus excluded from the values given for glaciers.
+10 100 Gt yr−1 of ice loss is equivalent to about 0.28 mm yr−1 of global mean sea level rise.
+
+SPMSummary for Policymakers101900 1920 1940 1960 1980 2000−20−1001020
+Year (1022 J)Change in global average upper ocean heat content (c)
+Global average sea level change
+1900 1920 1940 1960 1980 2000−50050100150200
+Year(mm)(d)Arctic summer sea ice extent
+1900 1920 1940 1960 1980 2000468101214
+Year(million km2)(b)Northern Hemisphere spring snow cover
+1900 1920 1940 1960 1980 200030354045
+Year(million km2)(a)
+Figure SPM.3 |  Multiple observed indicators of a changing global climate: (a) Extent of Northern Hemisphere March-April (spring) average snow cover; (b) 
+extent of Arctic July-August-September (summer) average sea ice; (c) change in global mean upper ocean (0–700 m) heat content aligned to 2006−2010, 
+and relative to the mean of all datasets for 1970; (d) global mean sea level relative to the 1900–1905 mean of the longest running dataset, and with all 
+datasets aligned to have the same value in 1993, the first year of satellite altimetry data. All time-series (coloured lines indicating different data sets) show 
+annual values, and where assessed, uncertainties are indicated by coloured shading. See Technical Summary Supplementary Material for a listing of the 
+datasets. {Figures 3.2, 3.13, 4.19, and 4.3; FAQ 2.1, Figure 2; Figure TS.1}
+
+SPM Summary for Policymakers11B.4 Sea Level
+The atmospheric concentrations of carbon dioxide, methane, and nitrous oxide have 
+increased to levels unprecedented in at least the last 800,000 years. Carbon dioxide 
+concentrations have increased by 40% since pre-industrial times, primarily from fossil fuel 
+emissions and secondarily from net land use change emissions. The ocean has absorbed 
+about 30% of the emitted anthropogenic carbon dioxide, causing ocean acidification (see 
+Figure SPM.4). {2.2, 3.8, 5.2, 6.2, 6.3}
+11 ppm (parts per million) or ppb (parts per billion, 1 billion = 1,000 million) is the ratio of the number of gas molecules to the total number of molecules of dry air. For example, 
+300 ppm means 300 molecules of a gas per million molecules of dry air.The rate of sea level rise since the mid-19th century has been larger than the mean rate 
+during the previous two millennia ( high confidence ). Over the period 1901 to 2010, global 
+mean sea level rose by 0.19 [0.17 to 0.21] m (see Figure SPM.3). {3.7, 5.6, 13.2}
+• Proxy and instrumental sea level data indicate a transition in the late 19th to the early 20th century from relatively low 
+mean rates of rise over the previous two millennia to higher rates of rise ( high confidence ). It is likely  that the rate of 
+global mean sea level rise has continued to increase since the early 20th century. {3.7, 5.6, 13.2}
+• It is very likely  that the mean rate of global averaged sea level rise was 1.7 [1.5 to 1.9] mm yr–1 between 1901 and 2010, 
+2.0 [1.7 to 2.3] mm yr–1 between 1971 and 2010, and 3.2 [2.8 to 3.6] mm yr–1 between 1993 and 2010. Tide-gauge and 
+satellite altimeter data are consistent regarding the higher rate of the latter period. It is likely  that similarly high rates 
+occurred between 1920 and 1950. {3.7}
+• Since the early 1970s, glacier mass loss and ocean thermal expansion from warming together explain about 75% of the 
+observed global mean sea level rise ( high confidence ). Over the period 1993 to 2010, global mean sea level rise is, with 
+high confidence , consistent with the sum of the observed contributions from ocean thermal expansion due to warming 
+(1.1 [0.8 to 1.4] mm yr–1), from changes in glaciers (0.76 [0.39 to 1.13] mm yr–1), Greenland ice sheet (0.33 [0.25 to 0.41] 
+mm yr–1), Antarctic ice sheet (0.27 [0.16 to 0.38] mm yr–1), and land water storage (0.38 [0.26 to 0.49] mm yr–1). The sum 
+of these contributions is 2.8 [2.3 to 3.4] mm yr–1. {13.3}
+• There is very high confidence  that maximum global mean sea level during the last interglacial period (129,000 to 116,000 
+years ago) was, for several thousand years, at least 5 m higher than present, and high confidence  that it did not exceed 
+10 m above present. During the last interglacial period, the Greenland ice sheet very likely  contributed between 1.4 and 
+4.3 m to the higher global mean sea level, implying with medium confidence  an additional contribution from the Antarctic 
+ice sheet. This change in sea level occurred in the context of different orbital forcing and with high-latitude surface 
+temperature, averaged over several thousand years, at least 2°C warmer than present ( high confidence ). {5.3, 5.6}
+B.5 Carbon and Other Biogeochemical Cycles
+• The atmospheric concentrations of the greenhouse gases carbon dioxide (CO2), methane (CH4), and nitrous oxide (N2O) 
+have all increased since 1750 due to human activity. In 2011 the concentrations of these greenhouse gases were 391 
+ppm11, 1803 ppb, and 324 ppb, and exceeded the pre-industrial levels by about 40%, 150%, and 20%, respectively. {2.2, 
+5.2, 6.1, 6.2}
+• Concentrations of CO2, CH4, and N2O now substantially exceed the highest concentrations recorded in ice cores during 
+the past 800,000 years. The mean rates of increase in atmospheric concentrations over the past century are, with very 
+high confidence , unprecedented in the last 22,000 years. {5.2, 6.1, 6.2}
+
+SPMSummary for Policymakers12• Annual CO2 emissions from fossil fuel combustion and cement  production were 8.3 [7.6 to 9.0] GtC12 yr–1 averaged over 
+2002–2011  (high confidence ) and were 9.5 [8.7 to 10.3] GtC yr–1 in 2011, 54% above the 1990 level. Annual net CO2 
+emissions from  anthropogenic land use change were 0.9 [0.1 to 1.7] GtC yr–1 on average during 2002 to 2011 ( medium 
+confidence ). {6.3}
+• From 1750 to 2011, CO2 emissions from fossil fuel combustion and cement production have released 375 [345 to 405] 
+GtC to the atmosphere, while deforestation and other land use change are estimated to have released 180 [100 to 260] 
+GtC. This results in cumulative anthropogenic emissions of 555 [470 to 640] GtC. {6.3}
+• Of these cumulative anthropogenic CO2 emissions, 240 [230 to 250] GtC have accumulated in the atmosphere, 155 [125 
+to 185] GtC have been taken up by the ocean and 160 [70 to 250] GtC have accumulated in natural terrestrial ecosystems 
+(i.e., the cumulative residual land sink). {Figure TS.4, 3.8, 6.3}
+• Ocean acidification is quantified by decreases in pH13. The pH of ocean surface water has decreased by 0.1 since the 
+beginning of the industrial era ( high confidence ), corresponding to a 26% increase in hydrogen ion concentration (see 
+Figure SPM.4). {3.8, Box 3.2}
+Figure SPM.4 |  Multiple observed indicators of a changing global carbon cycle: (a) atmospheric concentrations of carbon dioxide (CO2) from Mauna Loa 
+(19°32’N, 155°34’W – red) and South Pole (89°59’S, 24°48’W – black) since 1958; (b) partial pressure of dissolved CO2 at the ocean surface (blue curves) 
+and in situ pH (green curves), a measure of the acidity of ocean water. Measurements are from three stations from the Atlantic (29°10’N, 15°30’W – dark 
+blue/dark green; 31°40’N, 64°10’W – blue/green) and the Pacific Oceans (22°45’N, 158°00’W − light blue/light green). Full details of the datasets shown 
+here are provided in the underlying report and the Technical Summary Supplementary Material. {Figures 2.1 and 3.18; Figure TS.5}(a)
+(b)1950 1960 1970 1980 1990 2000 2010300320340360380400
+YearCO 2 (ppm) 
+1950 1960 1970 1980 1990 2000 2010320340360380400
+YearpCO 2 (μatm)
+8.068.098.12
+in situ pH unitSurface ocean CO 2 and pH  Atmospheric CO 2 
+12 1 Gigatonne of carbon = 1 GtC = 1015 grams of carbon. This corresponds to 3.667 GtCO2.
+13 pH is a measure of acidity using a logarithmic scale: a pH decrease of 1 unit corresponds to a 10-fold increase in hydrogen ion concentration, or acidity. 
+
+SPM Summary for Policymakers1314 The strength of drivers is quantified as Radiative Forcing (RF) in units watts per square metre (W m–2) as in previous IPCC assessments. RF is the change in energy flux 
+caused by a driver, and is calculated at the tropopause or at the top of the atmosphere. In the traditional RF concept employed in previous IPCC reports all surface and 
+tropospheric conditions are kept fixed. In calculations of RF for well-mixed greenhouse gases and aerosols in this report, physical variables, except for the ocean and sea 
+ice, are allowed to respond to perturbations with rapid adjustments. The resulting forcing is called Effective Radiative Forcing (ERF) in the underlying report. This change 
+reflects the scientific progress from previous assessments and results in a better indication of the eventual temperature response for these drivers. For all drivers other than 
+well-mixed greenhouse gases and aerosols, rapid adjustments are less well characterized and assumed to be small, and thus the traditional RF is used. {8.1}
+15 This approach was used to report RF in the AR4 Summary for Policymakers.Total radiative forcing is positive, and has led to an uptake of energy by the climate system. 
+The largest contribution to total radiative forcing is caused by the increase in the atmospheric 
+concentration of CO2 since 1750 (see Figure SPM.5). {3.2, Box 3.1, 8.3, 8.5}C. Drivers of Climate Change
+Natural and anthropogenic substances and processes that alter the Earth’s energy budget are drivers of climate change. 
+Radiative forcing14 (RF) quantifies the change in energy fluxes caused by changes in these drivers for 2011 relative to 1750, 
+unless otherwise indicated. Positive RF leads to surface warming, negative RF leads to surface cooling. RF is estimated based 
+on in-situ and remote observations, properties of greenhouse gases and aerosols, and calculations using numerical models 
+representing observed processes. Some emitted compounds affect the atmospheric concentration of other substances. The RF 
+can be reported based on the concentration changes of each substance15. Alternatively, the emission-based RF of a compound 
+can be reported, which provides a more direct link to human activities. It includes contributions from all substances affected 
+by that emission. The total anthropogenic RF of the two approaches are identical when considering all drivers. Though both 
+approaches are used in this Summary for Policymakers, emission-based RFs are emphasized.
+• The total anthropogenic RF for 2011 relative to 1750 is 2.29 [1.13 to 3.33] W m−2 (see Figure SPM.5), and it has increased 
+more rapidly since 1970 than during prior decades. The total anthropogenic RF best estimate for 2011 is 43% higher than 
+that reported in AR4 for the year 2005. This is caused by a combination of continued growth in most greenhouse gas 
+concentrations and improved estimates of RF by aerosols indicating a weaker net cooling effect (negative RF). {8.5}
+• The RF from emissions of well-mixed greenhouse gases (CO2, CH4, N2O, and Halocarbons) for 2011 relative to 1750 is 
+3.00 [2.22 to 3.78] W m–2 (see Figure SPM.5). The RF from changes in concentrations in these gases is 2.83 [2.26 to 3.40] 
+W m–2. {8.5}
+• Emissions of CO2 alone have caused an RF of 1.68 [1.33 to 2.03] W m–2 (see Figure SPM.5). Including emissions of other 
+carbon-containing gases, which also contributed to the increase in CO2 concentrations, the RF of CO2 is 1.82 [1.46 to 
+2.18] W m–2. {8.3, 8.5}
+• Emissions of CH4 alone have caused an RF of 0.97 [0.74 to 1.20] W  m−2 (see Figure SPM.5). This is much larger than the 
+concentration-based estimate of 0.48 [0.38 to 0.58] W m−2 (unchanged from AR4). This difference in estimates is caused 
+by concentration changes in ozone and stratospheric water vapour due to CH4 emissions and other emissions indirectly 
+affecting CH4. {8.3, 8.5}
+• Emissions of stratospheric ozone-depleting halocarbons have caused a net positive RF of 0.18 [0.01 to 0.35] W m−2 (see 
+Figure SPM.5). Their own positive RF has outweighed the negative RF from the ozone depletion that they have induced. 
+The positive RF from all halocarbons is similar to the value in AR4, with a reduced RF from CFCs but increases from many 
+of their substitutes. {8.3, 8.5}
+• Emissions of short-lived gases contribute to the total anthropogenic RF . Emissions of carbon monoxide (CO) are virtually 
+certain  to have induced a positive RF , while emissions of nitrogen oxides (NOx) are likely  to have induced a net negative 
+RF (see Figure SPM.5). {8.3, 8.5}
+• The RF of the total aerosol effect in the atmosphere, which includes cloud adjustments due to aerosols, is –0.9 [–1.9 to 
+−0.1] W m−2 (medium confidence ), and results from a negative forcing from most aerosols and a positive contribution 
+
+SPMSummary for Policymakers14from black carbon absorption of solar radiation. There is high confidence  that  aerosols and their interactions with clouds 
+have offset a substantial portion of global mean forcing from well-mixed greenhouse gases. They continue to contribute 
+the largest uncertainty to the total RF estimate. {7.5, 8.3, 8.5}
+• The forcing from stratospheric volcanic aerosols can have a large impact on the climate for some years after volcanic 
+eruptions. Several small eruptions have caused an RF of –0.11 [–0.15 to –0.08] W m–2 for the years 2008 to 2011, which 
+is approximately twice as strong as during the years 1999 to 2002. {8.4}
+• The RF due to changes in solar irradiance is estimated as 0.05 [0.00 to 0.10] W m−2 (see Figure SPM.5). Satellite obser -
+vations of total solar irradiance changes from 1978 to 2011 indicate that the last solar minimum was lower than the 
+previous two. This results in an RF of –0.04 [–0.08 to 0.00] W m–2 between the most recent minimum in 2008 and the 
+1986 minimum. {8.4}
+• The total natural RF from solar irradiance changes and stratospheric volcanic aerosols made only a small contribution to 
+the net radiative forcing throughout the last century, except for brief periods after large volcanic eruptions. {8.5}
+Figure SPM.5 |  Radiative forcing estimates in 2011 relative to 1750 and aggregated uncertainties for the main drivers of climate change. Values are 
+global average radiative forcing (RF14), partitioned according to the emitted compounds or processes that result in a combination of drivers. The best esti -
+mates of the net radiative forcing are shown as black diamonds with corresponding uncertainty intervals; the numerical values are provided on the right 
+of the figure, together with the confidence level in the net forcing (VH – very high , H – high, M – medium , L – low, VL – very low ). Albedo forcing due to 
+black carbon on snow and ice is included in the black carbon aerosol bar. Small forcings due to contrails (0.05 W m–2, including contrail induced cirrus), 
+and HFCs, PFCs and SF6 (total 0.03 W m–2) are not shown. Concentration-based RFs for gases can be obtained by summing the like-coloured bars. Volcanic 
+forcing is not included as its episodic nature makes is difficult to compare to other forcing mechanisms. Total anthropogenic radiative forcing is provided 
+for three different years relative to 1750. For further technical details, including uncertainty ranges associated with individual components and processes, 
+see the Technical Summary Supplementary Material. {8.5; Figures 8.14–8.18; Figures TS.6 and TS.7}
+Anthropogeni c Natural
+−1 0 1 2 3 
+ 
+ 
+Radiative forcing relative to 1750 (W m−2)Level of
+confidenceRadiative forcing by emissions and drivers
+1.68 [1.33 to 2.03] 
+0.97 [0.74 to 1.20]
+0.18 [0.01 to 0.35]
+0.17 [0.13 to 0.21]
+0.23 [0.16 to 0.30]
+0.10 [0.05 to 0.15]
+-0.15 [-0.34 to 0.03]
+-0.27 [-0.77 to 0.23]
+-0.55 [-1.33 to -0.06]
+-0.15 [-0.25 to -0.05]
+0.05 [0.00 to 0.10]
+2.29 [1.13 to 3.33]
+1.25 [0.64 to 1.86]
+0.57 [0.29 to 0.85]VH
+H
+H
+VH
+M
+M
+M
+H
+L
+M
+M
+H
+H
+MCO2
+CH4
+Halo-
+carbons
+N2O
+CO
+NMVOC
+NOxEmitted
+compound
+Aerosols and
+precursors
+(Mineral dust , 
+SO2, NH3,
+Organic carbon
+and Black carbon )Well-mixed greenhouse gases Short lived gases and aerosolsResulting atmospheric
+drivers
+CO2
+CO2H2OstrO3CH4
+O3CFCs HCFCs
+CO2CH4O3N2O
+CO2CH4O3
+Nitrate CH4O3
+Black carbonMineral dust
+Organic carbonNitrate Sulphate
+Cloud adjustments
+due to aerosols
+Albedo change
+due to land use
+Changes in
+solar irradiance
+Total anthropogenic
+RF relative to 1750
+195019802011
+

aaaifss2022_1.txt

@@ -0,0 +1,459 @@
+ AI-Based Text Analysis for Evaluati ng Food Waste Polic ies 
+John A. Aitken,1 Denali W. Rao, Balca Alaybek, Amber Sprenger, Grace Mika, Rob Hartman, 
+Laura Leets   
+The MITRE Corporation  
+[email protected]  
+ 
+ 
+ 
+Abstract  
+Food waste is a major contributor to climate change, making 
+the reduction of food waste one of the most important strate-
+gies to preserve threatened ecosystems and increase eco-
+nomic benefits. To evaluate the impact of food waste policies 
+in this arena and provide actionable guidance to policym ak-
+ers, we conducted  an AI -based  text analysis of  food waste 
+policy provisions . Specifically, we  used unsupervised ma-
+chine learning to  a) identif y commonalities across state pol-
+icy texts, b) cluster states by shared policy text, and c) exam-
+ine relationships  between  state cluster membership s and food 
+waste. This approach generated  state clusters  but demon-
+strate d very limited  convergent validity with policy ratings 
+provided by subject matter experts and no predictive validity 
+with food waste. We discuss the po tential of  using  supervised 
+machine learning to analyze food waste policy text as a next 
+step. 
+ 
+Keywords: food waste ; date label policy; text analysis  
+Introduction  
+Food waste is one of the most  significant driver s of climate 
+change, constituting  up to  10% of all greenhouse gas emis-
+sions , 14% of all water use , 18% of all cropland use , and 
+24% of all landfill content (Hall et al., 20 09; Quested, Ingle, 
+and Parry, 2013 ). The reduction of food waste is a key cli-
+mate change strategy (Hawke n, 2017),  and is  a challenge 
+that depends on a host of actors across all steps of the supply 
+chain. Strong  federal  and state policy is one of the most 
+promising avenues for mitigating food waste and stimulat-
+ing food recovery  (Evans and Nagele, 2018) . One key food 
+waste polic y in this area  and which is well-represented at the 
+state level is date label policy.   
+ Date label policy applies to  whether manufacturers must 
+include labels on certain food (e.g., milk, meat) and whether 
+the product may be sold past the date, and other  require-
+ments  such as the use of specific terminology (e.g., “Best 
+ 
+Copyright © 2022, Association for the Advancement of Artificial Intelli-
+gence (www.aaai.org). All rights reserved.  
+ by,” “Sell by,” “Use by”) . Currently , instead of a  federally -
+enforced  standard policy, there is a patchwork of date label 
+policies across states  that grants free reign to manufacturers 
+and, in turn, creates confusion for consumers (Broad Leib 
+and Pollans, 2019; Broad Leib et al., 2016) . For example, 
+consum ers may incorrectly believe that food should be dis-
+carded once past its date, whereas some dates may only be 
+indicators of quality rather than wholesomeness ( Busetti, 
+2019 ). It has been suggested that states with more extensive 
+date label policy (i.e., mor e requirements  and/or restrictions) 
+contribute to  rather than reduce food waste ( Lipinski et al., 
+2013; Povich, 20 20). However, there is a paucity of empiri-
+cal evidence regarding the impact of these policies on rele-
+vant outcomes.  
+ Therefore, it is essential to empirically evaluate  the extent 
+to which state date label  polic ies contribute to waste and im-
+pacts the environment. One challenge in accomplishing this 
+goal is that these policies can be generally opaque due to 
+legal jargon  and require expert analysis to distill. For exam-
+ple, beyond reading and comprehending a state’s date label 
+policy , an expert must also be able to evaluate the strengths 
+and weakness of the policy with respect to relevant policy at 
+the federal level and in ot her states as well as with respect 
+to the nuances of the issue at hand (i.e., knowing which pol-
+icy features are more or less beneficial in combating climate 
+change). In such an evaluation , subject matter experts may 
+differ in the extent to which they agree  about which policy 
+features  should be considered  (i.e., which are most relevant 
+to the efficacy of the policy) as well as how to judge  quali-
+tative aspects of the policy (e.g., strength, extensiveness) , 
+requiring prolonged discussions and recalibration . Such an 
+effort is extensive and time -consuming, which is far from 
+ideal given the significance of food waste in driving climate 
+change and the urgency for policymakers to craft effective 
+and relevant policies.  
+ 
+ 
+ 
+©2022 The MITRE Corporation. ALL RIGHTS RESERVED.Approved for  Public Release; Distribution Unlimited. Public Release Case Number 2 2-2533 
+ 
+
+ 2 
+  In this respect, text analysis with the aid of  artificial in-
+telligence (AI) represents a promising avenue of policy eval-
+uation in the food waste and climate change domains.  First, 
+AI-based text analysis may be an extremely efficient tool in 
+analyzing large quantities of policy text  and rendering a list 
+of key characteristics that differentiates one state’s policy 
+from another. This would greatly benefit domain experts 
+and researchers in general by functioning as a powerful and 
+flexible tool in many climate change policy areas ( Short , 
+McKenny, and Reid , 2018 ). For example, there may be a 
+high degree of shared text between policy texts  that may in-
+dicate  similarit ies between state policies . Also, there may be 
+unique features of certain policies differentiat ing them from  
+others . Overall, then, legislative text reuse and analysis 
+could serve  as a window into the spread of political influ-
+ence (Wilkerson, Smith , and Stramp , 2015 ).  
+ Moreover , developing an analytical method that can sum-
+marize and evaluate climate change  policy text may allow 
+for non -experts to investigate and interpret this policy area . 
+This enables a multidisciplinary approach to a typically 
+complex legislative area, and such an approach is crucial 
+given the scope of climate change and it s causes (of interest 
+to this paper, f ood waste) as well as  the variety of domains 
+(e.g., environmental  science , political  science , social sci-
+ence)  and stakeholders (e.g., federal agencies,  nonprofit or-
+ganizations)  involved in addressing these problems . 
+ To that end, t he natural language processing field com-
+bines AI and computational linguistic techniques and pro-
+vides a variety of machine learning approaches  (e.g., super-
+vised, unsupervised ) for text analysis . The rest of this paper 
+summarizes our application of unsuperv ised machine learn-
+ing to food waste  policy text analy sis. Specifically , we per-
+formed a text analysis of U.S. state food date label policies  
+to derive state clusters that (1) meaningfully represented the 
+content of shared policy text, (2) converged with hum an 
+subject -matter expert ratings of policies,  and (3) predicted 
+food waste. The content, convergent, and predictive v alida-
+tion of such a method would contribute to impact analyses 
+in not only date label policies but also other climate change 
+policy areas.  
+Method  
+The data for this study included (1) date label policy texts 
+from 50 U.S. states  enacted prior to 2012  and (2)  municipal 
+solid waste (MSW ; 22–24% of which is estimated to ac-
+count for food waste across states, U.S. Environmental Pro-
+tection Agency , 2022 ).  
+Policy Text Preparation  for Text Analysis  
+We processed the policy text at two levels of analysis : enti re 
+provision s (i.e., sections  with explicit citation labels)  and in-dividual clauses  within legislative provisions  (i.e., distin-
+guished by line breaks and enumeration marks ). We orga-
+nized  the data accordingly  and removed duplicate provi-
+sions , which resulted in  113 distinct provisions and 1846 
+distinct clauses in the  date label policy dataset.  We then to-
+kenized the text, using term frequency -inverse document 
+frequency  (TF-IDF) weighting  to create token -frequency 
+vector s. Finally, we discarded policy f ragments with fewer 
+than seven tokens  as they were too short to be meaningful .  
+Policy Text Coding for Validation  
+Food waste policy subject matter experts (SMEs)  from Har-
+vard Law School, Food Law and Policy Clinic  (HFLPC) 
+manually coded policy texts to generate  a ground -truth  char-
+acterization of the state date label polic ies against which the 
+text analysis -derived clusters  could be validated. We used 
+three manually coded variables for validation purposes : for 
+a given food type, (1) whether a date label i s required, (2) 
+whether sale after label date is restricted, and (3) whether 
+the policy required the use of specified date label terminol-
+ogy (e.g., “use by,” “sell by,” “best by”) . We created three 
+respective continuous variables (i.e., DateTotal, SaleTota l, 
+and TermTotal) that indicated the number of food types for 
+which a given date label policy was enacted in a given state.  
+AI-Based Text Analysis Plan  
+We took three steps in each policy text analysis. First, we 
+applied a standard topic modeling algorithm,  Gensim, to the 
+provision token frequency vectors . Given our set of non -uni-
+formly structured  text data, topic modeling was the natural 
+choice to begin processing and understanding the data.  Alt-
+hough  there were  8 distinct food types identified in our da-
+taset as policy foci ( Breads & Bakery, Dairy & Eggs, Dry 
+Goods, Fresh Meat & Seafood, Frozen, Pr epared Foods, 
+Produce, and Ready -to-drink Beverages ), it was important  
+to represent the data with more topics than just these 8  to 
+capture all possible fragments and more specific food types 
+(e.g., Shellfish within the  broader Fresh Meat & Seafood 
+category ). Additionally , while it is generally recommended 
+to run the Gensim topic modeling algorithm with 300 -500 
+topics, we determined that our dataset was unlikely to in-
+clude as many distinct topics ( Bradford , 2009 ). Preliminary 
+experimentation revealed that ex tracting more than 150 top-
+ics yielded many overlapping topics, while extracting fewer 
+than 80 topics from the dataset yielded topics that incorpo-
+rated unrelated concepts into one. Accordingly , we specified  
+the model to generate 100 topics . We represented t he policy 
+text fragments as proportions of the 100 topics and com-
+puted the cosine similarity between each pair of  policy frag-
+ment  topic vectors . We then applied a similarity threshold to 
+select only the stronger relationships between policy frag-
+ments  and generated a network graph to visualize the results, 
+
+ 3 
+plotting fragments as nodes and the relationship s between 
+them as edges.   
+ Second , we attemp ted to  group the policy fragments by 
+their semantic features.  We used agglomerative clustering  
+on the previously generated network graph  to detect groups 
+of similar policy fragments  and color -coded the graph to re-
+flect these clusters . Using a hierarchical clustering method  
+like agglomerative clustering allowed us to leave the num-
+ber of clusters to create unspecified and explore  how many 
+clusters “naturally” emerged from the data . The resulting  
+policy fragment  clusters  from this step , since they were 
+compute d via the application of  similarity metrics to topic 
+modeling  outputs,  represented  equivalence classes under 
+topic similarity .  
+ Third, we used the policy fragment clusters as features  
+and describe d each state as a combination of the features it 
+had. These descriptions took the form of feature vectors 
+(similar to the topic feature vectors we saw earlier, but with 
+one per state instead of per policy fragment). For example, 
+when  a state had one policy fragment that fell under  a dairy 
+labeling  cluster , had  two that fell under  the shellfish require-
+ments cluster, and had no policy about pork (meaning no 
+membership in a pork requirements cluster ), each of these 
+cluster memberships as well as non-memberships  were in-
+corporated in  the state’s feature vector . We then took the co-
+sine similarity of these state feature vectors and applied an-
+other similarity threshold to select only strong relationships  
+between states . From the resulting filtered stat e similarity 
+matrix, we generated a network graph using states as the 
+nodes in the graph and reflecting the strength of the similar-
+ity between them in the lengths of the edges. Finally, we de-
+tected clusters of states in the graph and color -coded those 
+clusters. We expected t hese state clusters to consist of  states 
+that were  similar to each other with respect to  their food date 
+label polic y content .  
+Results  
+Text  Reuse Analysis  
+We began by examining verbatim text reuse by employing 
+common subsequence analysis  to compute all common sub-
+sequences of at least 6 words between each pair of states.  
+We then used three analytical techniques : (1) extract ing the 
+length of the longest common subsequence between every 
+pair of states,  (2) comput ing the number of common subse-
+quences shared by each pair of states , and (3) extracting any 
+subsequences greater than 6 words long that were common 
+to more than two state s. However, this approach failed to 
+identify  identical provisions at the section level nor verbatim 
+duplication of meaningful policy expressions within provi-
+sions . Therefore, we instead shifted our approach to search-
+ing for similar  text between policies at  the level of the pro-
+vision and of individual clauses, hypothesizing that states with similar policy text share common policy objectives and 
+may be clustered as such.   
+Date Label Policy Text Analysis  
+Beginning with date label policy text at the provision -level, 
+we used the topic modeling algorithm to generate  100 top-
+ics. Figure 1 shows six of the twelve most significant topics’ 
+ten most strongly weighted tokens compared to the fre-
+quency of those tokens in the entirety of the text. Some top-
+ics were more clea rly interpreted than others: Topic 0  (shell-
+fish, tag, dealer, molluscan, shucked, shellstock, etc.)  clearly 
+revolve d around shellfish and how they should be caught 
+and processed; Topic 2 (egg, milk, carton, pack, size, inch, 
+etc.) seem ed to be about specifically egg cartons in contrast 
+to milk cartons; and Topic 10  (mean, sandwich, expiration, 
+prewrapped,  open, vendor, etc.)  seem ed to be about pre-
+wrapped sandwiches and their expiration dates. Each provi-
+sion was represented as a vector com bination of the 100 top-
+ics generated by our topic modeling algorithm.  
+We computed the pairwise cosine similarity between provi-
+sions based on this representation and applied a similarity 
+threshold of 0.6 (discarding any values below the similarity 
+threshol d). Agglomerative clustering (with distance thresh-
+old setting of 1.5) identified 20 clusters of provisions. We 
+used these clusters as features and represented the states as 
+vectors of length 20 denoting which clusters their provisions 
+fell into.  Most of th e clusters revolved around a certain food 
+type (milk, shellfish, prewrapped sandwiches) while a few 
+clusters were more general (pull dates, misbranding),  so the  
+number of features a state had was often a reflection of how 
+many different food types that state’s date label polic ies ad-
+dressed. The number of features a state  had was also partly 
+Figure 1. The t en most strongly weighted tokens from some of the top 
+twelve most significant topics  (Provision -Level Date Labeling Analysis)  
+
+ 4 
+a reflection of how many date -label -related provisions a 
+state had in total.  
+ Finally, we clustered the states themselves based on the 
+feature clusters  (see Figure 2) . We created a binary state fea-
+ture matrix, took the cosine similarity of the matrix, and dis-
+carded all values less than 0.5. Greedy modularity maximi-
+zation yi elded 8 clusters of states, with a modularity score 
+of 0.66.  
+ Turning now to the clause -level analyses,  many of the 
+same significant tokens emerged as in the provision -level 
+analysis, implying that the same tokens that were significant 
+within a whole prov ision are still the most significant when 
+the text is broken into smaller segments. On average, how-
+ever, it was harder to ascertain what these topics were  about. 
+Also notable  was that t he most significant tokens in these 
+topics were  less strongly weighted than the most significant 
+tokens in the provision -level topics. It is possible that the 
+fragmenting of the text to the clause level also split up im-
+portant or key phrases, so that indiv idually each clause had 
+fewer key phrases signaling its meaning.  
+ Agglomerative clustering (with a distance threshold set-
+ting of 4) detected 72 clusters of clauses. These clusters, as 
+may be expected, were more homogenous than the clusters 
+found at the pro vision -level analysis, both because the unit 
+of text was smaller and because there were more clusters for 
+them to separate into.  Greedy modularity maximization de-
+tected 6 clusters of states , with a modularity score of 0.69 .  
+Validation Analyses  
+ For conve rgent validat ion (i.e., examining the relation-
+ships between state s’ cluster  membership s and SME -coded  
+policy variables)  and predictive validation (i.e., examining 
+the relationships between state s’ cluster membership s and food waste), we first compute d two continuous s tate topic  
+count variable s (i.e., number of topics  within which a given 
+state fell under) . One of the continuous variables was based 
+on provisions and the other was based on clauses. Also, in 
+these variables, we included states that did not have any date 
+label policies, which received a value of zero. We computed 
+the Kendall’s Tau correlations of the state topic count vari-
+ables with the SME coded policy variables and the outcome 
+variable (MSW) , and found weak relationships  (τb = 0.19, p 
+= .079 for provisions; τb = 0.15, p = .133 for clauses ). More-
+over, after excluding the states with no date lab el policies 
+from the topic count variable, topic count and DateTotal var-
+iables, this correlation remained nonsignificant.  
+ In addition, we conducted a series of chi-squared differ-
+ence test s to examine whether state clusters were related to 
+SME -coded policy variables and MSW. These analyses re-
+turned nonsignificant results no matter states with no date 
+label policies included.  
+Discussion  
+Our findings general ly suggested that the unsupervised ma-
+chine learning approach for text analysis was able to cluster  
+food waste policy fragments and states based on similar fea-
+tures that emerged through the text, but the results demon-
+strated very limited  convergent validity  with those gener-
+ated by  SME  coding and no  predictive validity with the food 
+waste outcome . Our work in progress involves validating a 
+supervised machine learning approach to analyze policies 
+relevant to food waste and climate change.  
+ Additionally, future work could  perform more pre-pro-
+cessing of the policy text and employ more sophisticated 
+natural language proces sing ( NLP ) models . Although we 
+started our analyses by preparing  our text with standard  and 
+widely used text cleaning  methods , legal text often contains  
+additional levels of complexity (e.g., enumerations, hyper -
+specific abbreviations , particularly  formal phrasings ) com-
+pared to the type of text that our methods are commonly de-
+signed for and used on (e.g., social media posts, Wikipedia 
+articles).  Therefore, o ur dataset would likely benefit from 
+additional processing  that is more approp riate for policy 
+text. One potential  direction is to use a tool with pre -trained 
+word vectors  such as GloVe  (Pennington, Socher, and Man-
+ning, 2014) . Although it would be more computationally ex-
+pensive,  GloVe ’s incorporat ion of linguistic  and semantic 
+similari ty between words might be useful. For example, 
+equating the words “shellfish ” and “ mollusks ” might illumi-
+nate some previously hidden  policy  similarities in our da-
+taset. Another  potential  tool is LEGAL -BERT  (Chalkidis 
+et.al., 2020) . The authors  of LEGAL -BERT faced the same 
+issue we note above —that the usefulness of  standard pre-
+processing tools may not  generalize to legal tex t. We could 
+leverage t heir conclusion  (i.e., pre-training BERT models on 
+Figure 1. Color -coded state clust ers 
+ 
+Figure 2. Clause -Level Date Labeling Analysis : State Clusters  
+
+ 5 
+legal text improve s performance ) as well as their publ icly 
+released pre-trained  models  to improve our analyses .  
+ Moreover, the present work considered only policy text 
+related to food waste, which is an important arena in the cli-
+mate change discussion (Hall et al., 2009; Quested, Ingle, 
+and Parry, 2013 ), but future research may also consider leg-
+islative policy that is tied to other areas with a negative en-
+vironmental impact. We examined food waste as an initial 
+investigation and test of this methodology, and there is clear 
+potential for the examination of not only other policy texts 
+related to food waste (e.g., liability protection, tax incen-
+tives, etc.; Broad Leib et al., 2020) but related to sustainable 
+fishing (Worm et al., 2006)  and energy use (Hawken, 2017). 
+These other areas are of obvious relevanc e to climate change 
+and may include legislative policy that is amenable to such 
+analysis . 
+ Finally, a limitation of the current work is the reliance on 
+MSW as a proxy variable of food waste. Despite the likeli-
+hood that food waste exhibits significant conve rgence with 
+MSW, the limited predictive validity that we found in the 
+present work may be due to the MSW variable being  a 
+broader  measure that includes non -food related waste. Thus, 
+future work may evaluate food waste policies using a more 
+proximal or narr owly defined outcome variable. In general, 
+careful consideration of an appropriate outcome or indicator 
+variable is especially important in empirical evaluations of 
+legislative policy.  
+References  
+Bradford, R. 2008. An empirical study of required dimen-
+sionality for large -scale latent semantic indexing applica-
+tions. CIKM '08 : 153 -162.   
+Broad Leib, B.; Rice, C.; Neff, R.; Spiker, M.; Schklair, A.; 
+and Greenberg, S. 2016. Consumer Perceptions of Date La-
+bels: National Survey.  Safety , 23(54): 1 -4. 
+Broad Leib, E., and Pollans, M. J. 2019. The New Food 
+Safety. California Law Review, 107: 1173 -1248.  
+Broad Leib, E. ; Ardura, A. ; Fink, B. ; Hartman, M. ; Giguere, 
+M.; and Spiegler, R. 2020. United States Legal Guide: Food 
+Donation Law and Policy.  The Harvard  Law School Food 
+Law and Policy Clinic: Cambridge, MA, USA . Retrieved 
+from: https://chlpi.org/wp -content/uploads/2013/12/USA -
+Legal -Guide -2020.pdf  
+Busetti, S. 2019. A Theory -Based Evaluation of Food Waste 
+policy: Evidence from Italy.  Food Policy , 88: 101749 . 
+doi.org/10.1016/j.foodpol.2019.101749.  
+Chalkidis, I.; Fergadiotis,  M.; Malakasiotis, P.; Aletras , N.; 
+and Androutsopoulos , I. 2020.  LEGAL -BERT: The Mup-
+pets straight out of Law School. In Findings of Empirical 
+Methods in Natural Language Processing (EMNLP 2020)  
+https://aclanthology.org/2020.findings -emnlp.261  Evans, A. I.; and Nagele, R. M. 2018. A Lot to Digest: Ad-
+vancing Food Waste Policy in the United States.  Natural 
+Resou rces Journal , 58(1): 177 -214. jstor.org/sta-
+ble/26394778.  
+Hall, K. D.; Guo, J.; Dore, M.; and Chow, C. C. 2009. The 
+Progressive Increase of Food Waste in America and Its En-
+vironmental Impact.  PloS one , 4(11): e7940. 
+doi.org/10.1371/journal.pone.0007940.  
+Haw ken, P. 2017. Drawdown: The Most Comprehensive 
+Plan Ever Proposed to Reverse Global Warming.  New 
+York: Penguin Books.  
+Lipinski, B.; Hanson, C.; Waite, R.; Searchinger, T.; and Lo-
+max, J. 2013. Reducing Food Loss and Waste.  Washington: 
+World Resources Instit ute. 
+Pennington, J.; Socher, R.; and Manning, C. D. 20 14. 
+GloVe: Global Vectors for Word Representation . Proceed-
+ings of the 2014 Conference on Empirical Methods in Nat-
+ural Language Processing (EMNLP) . Association for Com-
+putational Linguistics .  
+Povich, E. S. 2019. Food Waste is a Major Problem. Con-
+fusing Date Labels are Making It Worse. Stateline . 
+Quested, T.; Ingle, R.; and Parry, A. 2013. Household Food 
+and Drink Waste in the United Kingdom 2012.  Banbury: 
+WRAP.  
+Short, J. C.; McKenny, A. F.; and Reid, S. W . 2018. More 
+than Words? Computer -Aided Text Analysis in Organiza-
+tional Behavior and Psychology Research.  Annual Review 
+of Organizational Psychology and Organizational Behav-
+ior, 5: 415 -435. doi.org/10.1146/annurev -orgpsych -
+032117 -104622.  
+U.S. Environmental  Protection Agency. 2022. Greenhouse 
+Gas Reporting Program . Retrieved from 
+https://www.epa.gov/ghgreporting.  
+Wilkerson, J.; Smith, D.; and Stramp, N. 2015. Tracing the 
+Flow of Policy Ideas in Legislatures: A Text Reuse Ap-
+proach.  American Journal of Politic al Science , 59(4): 943 -
+956. doi.org/10.1111/ajps.12175.  
+Worm, B. ; Barbier, E. B. ; Beaumont, N. ; Duffy, J. E. ; Folke, 
+C.; Halpern, B. S. ; ... and Watson, R. 2006. Impacts of Bio-
+diversity Loss on Ocean Ecosystem Services.  Sci-
+ence, 314(5800) : 787-790. doi.org/10.1126/sci-
+ence.1132294  
+Acknowledgments  
+This research was funded by the MITRE Independent Re-
+search and Development  Program.  
+ 
+The authors thank Charles A. Worrell for his constructive 
+feedback on earlier drafts of this paper.  
+

aaaifss2022_10.txt

@@ -0,0 +1,506 @@
+Contrastive Learning for Climate Model Bias Correction and Super-Resolution
+Tristan Ballard, Gopal Erinjippurath
+Sust Global
+San Francisco, California
+Abstract
+Climate models often require post-processing in order to
+make accurate estimates of local climate risk. The most
+common post-processing applied is bias-correction and spa-
+tial resolution enhancement. However, the statistical meth-
+ods typically used for this not only are incapable of captur-
+ing multivariate spatial correlation information but are also
+reliant on rich observational data often not available outside
+of developed countries, limiting their potential. Here we pro-
+pose an alternative approach to this challenge based on a
+combination of image super resolution (SR) and contrastive
+learning generative adversarial networks (GANs). We bench-
+mark performance against NASA’s flagship post-processed
+CMIP6 climate model product, NEX-GDDP. We find that our
+model successfully reaches a spatial resolution double that
+of NASA’s product while also achieving comparable or im-
+proved levels of bias correction in both daily precipitation
+and temperature. The resulting higher fidelity simulations of
+present and forward-looking climate can enable more local,
+accurate models of hazards like flooding, drought, and heat-
+waves.
+1 Introduction
+Global climate models by design are imperfect simulations
+of the physical world. While leading climate models like
+those in the Coupled Model Intercomparison Project phase
+6 (CMIP6) incorporate known phenomena like the laws of
+thermodynamics, other phenomena like cloud condensation
+have no known equations and require developers to include
+imprecise estimates. What’s more, climate models are run
+at spatial resolutions too coarse to simulate key phenom-
+ena like convective precipitation, tropical cyclone dynamics,
+and local effects from topography and land cover (microcli-
+mates). This leads to a variety of known and unknown er-
+rors, or biases, in projections of fundamental variables like
+temperature and precipitation.
+Climate model errors reduce the accuracy of projections
+of climate hazards like heatwaves and flooding, motivating
+the development of bias correction methods. These meth-
+ods (Section 2) generally involve deriving correction fac-
+tors to better align modeled historical values with observed
+historical values. The correction factors are then applied
+Copyright © 2022, Association for the Advancement of Artificial
+Intelligence (www.aaai.org). All rights reserved.
+ClimaGANRaw CMIP6
+Raw CMIP6
+ClimaGAN
+Precipitation [mm/day]
+Temperature [°C]Figure 1: Application of the ClimaGAN network to a CMIP6
+test set image (May 18, 1994) yields bias-corrected and 4x
+(0.5◦→0.125◦) super-resolution outputs. The network can
+be applied to CMIP6 daily simulations out to 2100.
+to forward-looking modeled values and are widely imple-
+mented in the climate impacts community. Indeed, forward-
+looking estimates of future flood risk typically use bias-
+corrected precipitation rather than the raw climate model
+data [1]. To enable local, accurate hazard models requires
+high fidelity, bias-corrected simulations of present-day and
+forward looking fundamental variables.
+Recent advances in AI including in image super-
+resolution (SR) and unpaired image-to-image translation
+suggest substantial promise to improve over existing bias
+correction methods. These AI models can flexibly incorpo-
+rate multivariate and spatial relationships in ways not pos-
+sible with existing approaches. For instance, AI-based SR
+has shown superior performance in enhancing the spatial
+resolution of wildfires [2], precipitation [3, 4, 5], and wind
+[6, 7]. Meanwhile, unpaired generative adversarial networks
+(GANs) have shown promise in applications to temperature
+[8] and precipitation [8, 9].
+Here we propose ClimaGAN, a novel SR and un-
+paired image-to-image translation GAN architecture oper-
+ating on 3-channel geospatial datasets, incorporating tem-
+
+perature, precipitation, and elevation. We validate and com-
+pare ClimaGAN performance against a NASA benchmark
+algorithm, showcasing ClimaGAN performance on a lead-
+ing CMIP6 model over a region spanning the contiguous
+U.S.
+2 Related Work
+There are several methods for bias-correcting and resolu-
+tion enhancement (downscaling) of climate variables, but
+the predominant method implemented in the climate com-
+munity is the bias-correction spatial disaggregation (BCSD)
+algorithm. For example, BCSD, proposed in 2002 [10], is
+the method used for NASA’s flagship CMIP6 bias-corrected
+product [11](Section 3.4). The bias-correction portion of
+BCSD is achieved through simple quantile mapping be-
+tween modeled and observed cumulative distribution func-
+tions. The resolution enhancement is achieved through ap-
+plication of Fourier transforms.
+The chief limitation of the (BCSD) algorithm used in
+NASA’s NEX-GDDP product [11] is that it is a simple sta-
+tistical method incapable of incorporating auxiliary datasets
+or spatial variability. For example, the only data that can
+be used to bias correct a modeled temperature dataset us-
+ing BCSD is an observed temperature dataset. However, we
+know that temperature biases are tightly linked to local fea-
+tures like elevation [12]. BCSD also implements bias cor-
+rection independently for each pixel, ignoring spatial cor-
+relation structure that can provide useful signal for further
+reducing biases.
+BCSD also does not permit multivariate relationships be-
+tween climate variables, despite the fact that most climate
+variables covary. Bias correcting temperature independently
+from precipitation, for example, can inadvertently introduce
+unrealistic relationships, particularly for extremes [13]. Bi-
+variate BCSD has been proposed but has not been widely
+adopted [13].
+We are aware of two recent AI-based approaches for cli-
+mate model bias correction, but neither incorporate spatial
+resolution enhancement. Both approaches are based on un-
+paired image-to-image translation, with one adapting the cy-
+cleGAN framework [9] and the other using UNIT [8]. Ex-
+treme learning machines have also been proposed as an al-
+ternative to BCSD [14].
+3 Data
+Geospatial data coded as input images to the model archi-
+tecture (Fig. 2) have 3 channels, corresponding to maps of
+daily temperature, daily precipitation, and elevation. Low
+resolution (LR) input images come from CMIP6 climate
+model simulations regridded to a common 0.5° resolution,
+while high resolution (HR) input images come from ob-
+served weather data regridded to a common 0.125° (14km)
+resolution.
+The study area covers the contiguous U.S., southern
+Canada, and northern Mexico, spanning 23°N and 49°N
+and 125°W and 65°W (Fig. 1). The LR input images are
+of dimension 54x120x3 while HR images are of dimension
+216x480x3.We train the model on 24 years of data from 1985 to 2014,
+setting aside 6 years [1990, 1994, 2000, 2004, 2008, 2012]
+of data in that period for testing. This results in 8,756 daily
+images for training and 2,194 daily images for testing.
+3.1 CMIP6 climate model simulations
+We demonstrate the ClimaGAN network with the U.S. Na-
+tional Oceanic and Atmospheric Administration’s Geophys-
+ical Fluid Dynamics Laboratory model GFDL-CM4, a lead-
+ing CMIP6 model [15]. The ClimaGAN network can be re-
+trained and applied to any of the CMIP6 models, an exten-
+sion of the current research we are actively pursuing (Sec-
+tion 7). The network can also be applied to corresponding
+CMIP6 forward looking (2015-2100) projections to derive
+estimates of future hazards (not shown).
+While CMIP6 models simulate a range of climate vari-
+ables, we focus here on simulations of daily maximum
+temperature and daily precipitation because these are often
+needed to derive climate hazards. The historical CMIP6 sim-
+ulations incorporate known values of carbon emissions, so-
+lar activity, and volcanic eruptions, among other inputs.
+3.2 Elevation
+We incorporate elevation data from the National Center for
+Atmospheric Research [16] as a supplementary feature to
+inform bias correction. Elevation is an important driver of
+local climate, so we expect it to be informative in bias cor-
+recting both temperature and precipitation.
+3.3 Observations
+We use the European Centre for Medium-Range Weather
+Forecasts (ECMWF) ERA5-Land data for observed daily
+maximum temperature and daily precipitation [17]. The re-
+analysis data has global coverage at approximately 9km res-
+olution over land, which we regrid to a coarser 0.125° for
+the analysis.
+3.4 NASA NEX-GDDP benchmark product
+We benchmark model performance against NASA’s flagship
+CMIP6 bias-corrected product, NEX-GDDP [11]. NEX-
+GDDP is based on the BCSD algorithm (Section 2). We
+use the daily maximum temperature and precipitation NEX-
+GDDP data corresponding to the same GFDL-CM4 model,
+such that outputs between the two methods are directly com-
+parable. NEX-GDDP data is available at 0.25° resolution.
+Beyond the technical limitations of NEX-GDDP (Section
+2), practical limitations for users are that it is not updated
+with the latest observational datasets and covers only a few
+variables and climate scenarios. The amount of available ob-
+servational data is projected to increase substantially with
+the release of new satellite and sensor datasets, yet NEX-
+GDDP will not begin to incorporate that new data until the
+release of CMIP7 years from now, if at all. This means that
+any advances in monitoring in data-poor regions, such as in
+many developing countries, will not be incorporated. Fur-
+ther, NEX-GDDP only covers 9 climate variables from the
+ScenarioMIP project, despite there being hundreds of other
+variables and MIP projects within CMIP6 of interest to re-
+searchers, limiting its scope.
+
+22CMIP6Daily imageGenerator
+DiscriminatorObservations(ERA5)BiasCorrectedCMIP6TemperaturePrecipitationElevationSuper-Resolution(4x)0.5°0.125°
+0.125°Figure 2: The ClimaGAN network takes as input daily CMIP6 climate data, as well as supplementary features like elevation,
+and outputs corresponding high-resolution, bias-corrected daily data. The network combines two key modules: super-resolution
+(SR) and a contrastive unpaired translation GAN. The SR layers enhance spatial resolution by 4x (0.5◦→0.125◦), while the
+GAN iteratively learns to bias-correct climate model inputs from comparisons with real-world observations.
+4 Methodology
+4.1 ClimaGAN Architecture
+We identified four key design goals for our network archi-
+tecture:
+• Unpaired image-to-image translation
+• Content preservation
+• Spatial resolution enhancement (super-resolution)
+• Multivariate input and output variables
+Unpaired image-to-image translation is required because
+the daily output from a CMIP6 model is not expected to
+directly match observations for the corresponding date, a
+challenge for typical bias-correction methods. For example,
+CMIP6 temperature simulations for Jan 1, 2010 are not, by
+design, expected to match observed conditions on that date.
+They instead are expected to provide a realistic simulation
+of what the weather could have been on that date.
+Content preservation is the idea that the bias-corrected
+output variables should maintain the content of the CMIP6
+inputs while taking on the appearance of real-world condi-
+tions. Content preservation in the context of GANs is typi-
+cally preserved through adding a cycle-consistency loss term
+[18].
+To achieve these design goals, we designed a network
+(ClimaGAN) that combines super-resolution and a con-
+trastive unpaired translation GAN (Fig. 2). The input LR im-
+ages passed through the network first go through two SR lay-
+ers that enhance spatial resolution by 4x. These SR imagesare then passed through a generator network. The discrimi-
+nator compares the output images with observation images
+to determine which image is ’real’ (observation) and which
+image is ’fake’ (bias-corrected and super-resolved CMIP6).
+The generator and discriminator networks along with the
+super-resolution layers are trained concurrently. As the gen-
+erator and discriminator improve, so does the level of bias-
+correction, creating output climate images that are increas-
+ingly difficult to distinguish from real-world observations.
+The generator consists of 9 Resnet blocks in between two
+upscaling layers (‘encoder’) and two downscaling layers
+(‘decoder’). The discriminator consists of 3 convolutional
+layers. The network contains approximately 14M parame-
+ters, and we train it for 20 epochs on an NVIDIA Tesla A100
+GPU.
+One of the key advances of this network is the imple-
+mentation of a contrastive unpaired translation GAN. The
+contrastive unpaired translation is an advancement in GANs
+released in 2020 from the team who created cycleGAN, a
+leading framework for unpaired image-to-image translation
+[19]. Contrastive unpaired translation appears to outperform
+cycleGANs in both accuracy and efficiency [19]. Briefly,
+the network incorporates a InfoNCE loss term [20] in ad-
+dition to the adversarial loss of a standard GAN [19]. The
+InfoNCE loss works by sampling patches of the output im-
+age and ensuring that the samples are similar to the corre-
+sponding patches of the input image. At the same time, the
+InfoNCE loss discourages the sampled patches from being
+too similar to other patches of the input image. This loss
+
+TemperaturePrecipitation
+ObservationClimaGAN (ours)NASA NEX-GDDPRaw CMIP6
+Temperature [°C]Precipitation [mm/day]
+Figure 3: ClimaGAN quantitatively and qualitatively enhances CMIP6 mean daily temperature (top) and daily precipitation
+(bottom) in a held out test set ( n=2,194 daily images). The scatter plots show the pixel-by-pixel correlations against observations
+across the U.S., while the maps show the southwestern U.S.
+term achieves content preservation. Further details on In-
+foNCE loss and the corresponding architecture additions can
+be found in Park et al.[19] We use the same default hyper-
+parameters for the number and size of patches as in Park et
+al. [19]
+4.2 Validation
+To validate the model, we measure correspondence between
+observations and ClimaGAN output on a held out test set
+(Section 3). We selected four statistical measures to as-
+sess the fidelity of model simulations in representing the
+observed statistical distribution: mean, standard deviation,
+skew, and the 98thpercentile. The 98thpercentile reflects the
+ability of the model to capture extremes. These statistics are
+computed for each pixel and then plotted as maps (Fig. 3) or
+aggregated across pixels using R2(Tables 1 and 2).
+5 Results
+We find that ClimaGAN substantially improves CMIP6 in-
+put simulations of daily temperature and precipitation, not
+only enhancing spatial resolution 4x to 0.125◦but also lead-
+ing to reductions in bias when evaluated on the held out test
+set.
+We first evaluate performance enhancement qualitatively
+by comparing maps of observed conditions against modeled
+(Fig. 3). Figure 3 shows mean conditions over the collection
+of daily test set images. Visually, the ClimaGAN-enhanced
+CMIP6 conditions much better match observed compared
+with the raw CMIP6 input, capturing local spatial variability
+with higher accuracy. In California, the Central Valley is re-
+flected clearly in enhanced temperatures, while the eastward
+Sierra Nevada mountains are reflected by a band of elevated
+precipitation, distinctions not immediately apparent in the
+original CMIP6 data (Fig. 3).Temperature Mean SD Skew Q98
+ClimaGAN (ours) 0.98 0.97 0.26 0.94
+NASA NEX-GDDP 0.96 0.90 0.69 0.86
+Raw CMIP6 0.94 0.88 0.42 0.75
+Table 1: ClimaGAN applied to daily maximum temperature
+shows an enhancement of raw CMIP6 inputs in out of sam-
+ple test set years ( n=2,194 daily images) across all four eval-
+uation metrics over the U.S. and outperforms NASA’s prod-
+uct except for distribution skew. Q98 = 98thpercentile.
+Precipitation Mean SD Skew Q98
+ClimaGAN (ours) 0.85 0.80 0.39 0.78
+NASA NEX-GDDP 0.86 0.81 0.42 0.80
+Raw CMIP6 0.78 0.69 0.37 0.72
+Table 2: ClimaGAN applied to daily precipitation shows an
+enhancement of raw CMIP6 inputs in out of sample test set
+years ( n=2,194 daily images) across all four evaluation met-
+rics over the U.S., though NASA’s product slightly outper-
+forms ClimaGAN. Q98 = 98thpercentile.
+Next, we evaluate performance enhancement quantita-
+tively in the held out test set, finding ClimaGAN improves
+over raw CMIP6 data across all four statistical measures for
+precipitation and across three of four statistical measures for
+temperature (Tables 1 and 2). For temperature, mean daily
+temperature improves from an R2of 94% with the original
+CMIP6 data to an R2of 98% after applying ClimaGAN (Fig.
+3; Table 1). We also find that extreme temperature, repre-
+sented by the 98thpercentile, improves from an R2of 75%
+to an R2of 94% (Table 1). Likewise for precipitation, mean
+
+daily precipitation improves from an R2of 78% with the
+original data to an R2of 85% after applying ClimaGAN
+(Fig. 3; Table 2). We also find that extreme precipitation,
+represented by the 98thpercentile, improves from an R2of
+72% to an R2of 78% (Table 2). The weaker performance of
+ClimaGAN on distributional skew suggests improvements
+can be made in capturing aspects of extremes, with one po-
+tential cause we are exploring further being the initial nor-
+malization steps applied to the data inputs.
+While these performance enhancements from ClimaGAN
+are promising, we are next curious how they compare
+against enhancements from a benchmark product, NASA’s
+NEX-GDDP bias corrected dataset.
+Benchmarking ClimaGAN performance against NASA’s
+product, the first key qualitative distinction is that Clima-
+GAN outperforms NASA’s product in capturing local spatial
+variability (Fig. 3). This is because ClimaGAN implements
+SR to twice the spatial resolution (0.125°) of NASA’s prod-
+uct (0.25°). Visually, the ClimaGAN-enhanced CMIP6 con-
+ditions better match observations compared with NASA’s
+product for temperature (Fig. 3). For precipitation, NASA’s
+product appears to better match observations in some areas,
+in part because it has less spatial variability than ClimaGAN
+(Fig. 3).
+Benchmarking performance quantitatively, we find that
+ClimaGAN leads to comparable or improved levels of bias
+correction as NASA’s product. For temperature, ClimaGAN
+outperforms NASA’s product on 3 of 4 metrics consid-
+ered, failing to improve the distributional skew metric (Table
+1). Particularly promising is that ClimaGAN yields an R2
+of 94% for extreme 98thpercentile temperature, compared
+with 86% for NASA’s product (Table 1). For precipitation,
+NASA’s product outererforms ClimaGAN on all 4 metrics,
+but the diffrences in performance are small, with R2differ-
+ences ranging from only 1-3% (Table 2).
+6 Conclusion
+Here we propose a framework for bias correcting and super-
+resolving daily climate model inputs to enable more accu-
+rate and high spatial resolution simulations of present day
+and future risk. The framework has several key advantages
+compared to other commonly employed approaches. First,
+it allows for superior levels of data-driven spatial resolution
+enhancement using super-resolution techniques. Second, it
+jointly bias corrects climate variables, allowing the model
+to learn from the multivariate relationship between climate
+variables and more accurately represent multivariate hazards
+like drought. Third, it flexibly incorporates additional geo-
+science variables like elevation to inform bias-correction.
+We find that ClimaGAN yields comparable or improved
+levels of bias-correction at twice the spatial resolution
+(14km) of NASA’s leading product (25km). This is excit-
+ing in part because there are numerous modifications to the
+network possible that could improve performance (Section
+7), while NASA’s product can only improve as the qual-
+ity of observational data improves. Moreover, NASA typi-
+cally releases and updates their product once every several
+years, while ClimaGAN can be regularly updated with thelatest sources of data. We expect this ability to update Clima-
+GAN with the latest observational data as it comes online
+as well as learn relationships in data-rich regions will help
+improve bias correction and SR in historically data-poor re-
+gions, such as in many developing countries (Section 7).
+Validation results for ClimaGAN suggest substantial po-
+tential for high resolution, enhanced accuracy projections of
+climate risk. Improvements in spatial resolution are critical
+to capturing local, asset-level effects of climate hazards. We
+found bias-correction improvements to the raw input data
+across metrics (Tables 1 and 2), and we highlight that the im-
+provements for extreme precipitation and extreme tempera-
+ture will enable more accurate projections of hazards like
+heatwaves and inland flooding. The higher fidelity simula-
+tions of present and forward-looking climate variables made
+possible by applying ClimaGAN can enable more local, ac-
+curate models of climate hazards, supporting climate scien-
+tists and a broad range of stakeholders alike.
+7 Future Directions
+We see several avenues for expanding and improving the
+ClimaGAN modeling approach. First, we intend to incor-
+porate additional global regions and CMIP6 models. Sec-
+ond, because ClimaGAN can flexibly integrate additional
+input channels, we can include variables like humidity, pres-
+sure, and wind to not only bias correct those variables but
+also improve accuracy on the temperature and precipitation
+variables. Architecture modifications to account for the spar-
+sity of precipitation may further improve results, including
+distributional skew [21]. Last, while we focus here on sin-
+gle image bias correction and SR, we see opportunities for
+improved performance on day-to-day variability by using
+multi-temporal images, which has been applied to satellite
+imagery in the past [22] but never, to our knowledge, to cli-
+mate model maps.
+References
+[1] Salman, A.M. and Y . Li. Flood risk assessment, future
+trend modeling, and risk communication: a review of
+ongoing research. Nat. Hazards Rev , 19(3), 2018.
+[2] Ballard, T. and G. Erinjippurath. FireSRnet:
+Geoscience-Driven Super-Resolution of Future
+Fire Risk from Climate Change. arXiv preprint
+arXiv:2011.12353 , 2020.
+[3] Jianxin Cheng, Qiuming Kuang, Chenkai Shen, Jin
+Liu, Xicheng Tan, and Wang Liu. Reslap: Generat-
+ing high-resolution climate prediction through image
+super-resolution. IEEE Access , 8:39623–39634, 2020.
+[4] Vandal, T., E. Kodra, S. Ganguly, and A. Michaelis.
+Deepsd: Generating high resolution climate change
+projections through single image super-resolution.
+Proceedings of the 23rd acm sigkdd international
+conference on knowledge discovery and data mining ,
+2017.
+[5] Vaughan, A., N.D. Lane, and M. Herzog. Multi-
+variate climate downscaling with latent neural pro-
+cesses. Tackling Climate Change with Machine Learn-
+ing ICML Workshop , 2021.
+
+[6] Kurinchi-Vendhan, R., B. Lutjens, R. Gupta, L.
+Werner, and D. Newman. WiSoSuper: Benchmarking
+Super-Resolution Methods on Wind and Solar Data.
+arXiv preprint arXiv:2109.08770 , 2021.
+[7] Stengel, K., A. Glaws, D. Hettinger, and R.N. King.
+Adversarial super-resolution of climatological wind
+and solar data. Proceedings of the National Academy
+of Sciences , 117(29), 2020.
+[8] Fulton, J.D. and B.J. Clarke. Towards debiasing cli-
+mate simulations using unsupervised image-to-image
+translation networks. Tackling Climate Change with
+Machine Learning NeurIPS Workshop , 2021.
+[9] Pan, B., G.J. Anderson, A. Goncalves, D.D. Lucas,
+C.J.W. Bonfils, J. Lee, Y . Tian, and H. Ma. Learn-
+ing to Correct Climate Projection Biases. Journal of
+Advances in Modeling Earth Systems , 13(10), 2021.
+[10] Wood, A.W., E.P. Maurer, A. Kumar, and D.P. Letten-
+maier. Long-range experimental hydrologic forecast-
+ing for the eastern United States. Journal of Geophys-
+ical Research Atmospheres , 107(D20), 2002.
+[11] Thrasher, B., W. Wange, A. Michaelis, F. Melton, T.
+Lee, and R. Nemani. NASA Global Daily Downscaled
+Projections, CMIP6. Scientific Data , 9(1), 2022.
+[12] Y . Lun, L. Liu, L. Cheng, X. Li, H. Li, and Z. Xu. As-
+sessment of GCMs simulation performance for precip-
+itation and temperature from CMIP5 to CMIP6 over
+the Tibetan Plateau. International Journal of Clima-
+tology , 41(7), 2021.
+[13] Li, C., E. Sinha, D.E. Horton, N.S. Diffenbaugh, and
+A.M. Michalak. Joint bias correction of temperature
+and precipitation in climate model simulations. Jour-
+nal of Geophysical Research: Atmospheres , 119(23),
+2014.
+[14] Zhang, S., F. Chen, X. He, and B. Liu. A new down-
+scaling approach and its performance with bias correc-
+tion and spatial disaggregation as contrast. Journal of
+Water and Climate Change , 8(4), 2017.
+[15] Held, I., H. Guo, A. Adcroft, J.P. Dunne, L.W.
+Horowitz, J. Krasting, E. Shevliakova, M. Winton, M.
+Zhao, M. Bushuk, and A.T. Wittenberg. Structure and
+performance of GFDL’s CM4.0 climate model. Jour-
+nal of Advances in Modeling Earth Systems , 11(11),
+2019.
+[16] National Geophysical Data Center/NESDIS-
+/NOAA/U.S. Department of Commerce. TerrainBase,
+Global 5 Arc-minute Ocean Depth and Land Elevation
+from the US National Geophysical Data Center
+(NGDC).
+[17] Munoz-Sabater, J., E. Dutra, A. Agusti-Panareda,
+C. Albergel, G. Arduini, G. Balsamo, S. Bous-
+setta, M. Choulga, S. Harrigan, H. Hersbach, and B.
+Martens. ERA5-Land: A state-of-the-art global reanal-
+ysis dataset for land applications. Earth System Sci-
+ence Data , 13(9), 2021.
+[18] Zhu, J.Y ., T. Park, P. Isola, and A.A. Efros. Unpaired
+image-to-image translation using cycle-consistent ad-versarial networks. Proceedings of the IEEE inter-
+national conference on computer vision , 2223-2232,
+2017.
+[19] Park, T., A.A. Efros, R. Zhang, and J.Y . Zhu. Con-
+trastive learning for unpaired image-to-image transla-
+tion. European conference on computer vision , 319-
+345, 2020.
+[20] Oord, A. Y . Li, and O. Vinyals. Representation learn-
+ing with contrastive predictive coding. arXiv preprint
+arXiv:1807.03748 , 2018.
+[21] Pathak, J., S. Subramanian, P. Harrington, S. Raja, A.
+Chattopadhyay, M. Mardani, T. Kurth, D. Hall, Z. Li,
+Z. Azizzadenesheli, and P. Hassanzadeh. Fourcastnet:
+A global data-driven high-resolution weather model
+using adaptive fourier neural operators. arXiv preprint
+arXiv:2202.11214 , 2022.
+[22] Deudon, M., A. Kalaitzis, I. Goytom, Md.R. Arefin,
+Z. Lin, K. Sankaran, V . Michalski, S.E. Kahou, J.
+Cornebise, and Y . Bengio. Highres-net: Recursive fu-
+sion for multi-frame super-resolution of satellite im-
+agery. arXiv preprint arXiv:2002.06460 , 2020.
+

aaaifss2022_11.txt

@@ -0,0 +1,460 @@
+Employing Deep Learning to Quantify Power Plant Greenhouse Gas Emissions via
+Remote Sensing Data
+Aryan Jain
+Amador Valley High School
+1155 Santa Rita Rd, Pleasanton, CA 94566
+Pleasanton, California 94588
+[email protected]
+Abstract
+Greenhouse gasses (GHG) emitted from fossil-fuel-burning
+power plants pose a global threat to climate and public
+health. GHG emissions degrade air quality and increase the
+frequency of natural disasters five-fold, causing 8.7 million
+deaths per year. Quantifying GHG emissions is crucial for
+the success of the planet. However, current methods to track
+emissions cost upwards of $520,000/plant. These methods are
+cost prohibitive for developing countries, and are not globally
+standardized, leading to inaccurate emissions reports from
+nations and companies. I developed a low-cost solution via an
+end-to-end deep learning pipeline that utilizes observations of
+emitted smoke plumes in satellite imagery to provide an accu-
+rate, precise system for quantifying power plant GHG emis-
+sions by segmentation of power plant smoke plumes, classi-
+fication of the plant fossil fuels, and algorithmic prediction
+of power generation and CO 2emissions. The pipeline was
+able to achieve a segmentation Intersection Over Union (IoU)
+score of 0.841, fuel classification accuracy of 92%, and quan-
+tify power generation and CO 2emission rates with R2val-
+ues of 0.852 and 0.824 respectively. The results of this work
+serve as a step toward the low-cost monitoring and detection
+of major sources of GHG emissions, helping limit their catas-
+trophic impacts on climate and our planet.
+Introduction
+Fossil-fuel power plants are one of the largest emitters of
+Greenhouse gasses, accounting for 73% of the U.S.’ GHG
+emissions and 65% of global GHG emissions (on Cli-
+mate Change and Edenhofer 2014). The pollutants produced
+by these emissions serve as major contributors to the climate
+crisis and have had devastating impacts on air quality and
+the environment. GHG emissions cause 8.7 million deaths
+per year and have increased the frequency of natural disas-
+ters such as wildfires and powerful storms five fold (Smol
+2012). These public health and environmental impacts cost
+billions in annual damages.
+Preventing the permanent effects of climate change and
+air pollution requires identifying the sources and distribu-
+tions of GHG emissions on a precise scale. However, keep-
+ing track of GHG emissions from all global power plants is
+difficult, as the quality of emissions data varies depending on
+Copyright © 2022, Association for the Advancement of Artificial
+Intelligence (www.aaai.org). All rights reserved.each country’s reporting protocols, maturity of their infras-
+tructure, and availability of proper monitoring systems. For
+example, in a developed country such as the United States,
+every major power plant has on-site Continuous Emissions
+Monitoring Systems (CEMS) that reports data to the Envi-
+ronmental Protection Agency. But these systems are very ex-
+pensive, costing over $500,000 for installation and $20,000
+annually for maintenance (US EPA 2016), making them im-
+practical for use in many lesser-developed countries. Ad-
+ditionally, the lack of reliable infrastructure causes many
+countries to provide vague, inaccurate, and outdated esti-
+mations of their GHG emissions. An examination of GHG
+emission reports from 196 countries found gaps in nation’s
+declared emissions versus estimates by the United Nations
+totalling to 10.6 billion tons of globally under-reported emis-
+sions per year (Mooney et al.).
+These issues require new, low-cost alternatives to estimate
+and report GHG emissions on a more precise scale. In recent
+years, the use of satellite data has emerged as a potential
+candidate to monitor the progression of climate change and
+global warming (Boesch et al. 2021). Equipped with an array
+of sensors and instruments to measure various atmospheric
+conditions, spectrometer satellites have helped inform our
+understanding of the dynamics of changes in Earth’s tem-
+perature. Launched in 2009 and 2014, spectrometer satellite
+missions Greenhouse Gasses Observing Satellite (GOSAT)
+and Orbiting Carbon Observatory (OCO-2) have provided
+carbon dioxide (CO 2) emission data on a global and national
+level (Eldering et al. 2017). However, spectrometer satellite
+instruments are imprecise and low-resolution ( ≥10 km res-
+olution), and cannot identify the granular emissions ( ≤2km)
+of individual power plants (Apte et al. 2017).
+When active, fossil-fuel burning power plants emit a
+smoke plume as a byproduct of the electricity generation
+process. These plumes can be captured by optical satellite
+imagery and fed into a deep learning model to produce accu-
+rate estimates of the plant’s GHG emissions (Cusworth et al.
+2021). Pairing deep learning with high-resolution optical
+satellite imagery serves as a promising method to estimate
+power plant GHG emissions with accuracy rates near spec-
+trometer measurements, while simultaneously maintaining
+the ability to monitor emissions on a global scale. More-
+over, this method does not require huge investments or elab-
+orate infrastructure, serving as a low-cost alternative to fill-
+
+ing long existing gaps in emissions data around the world.
+In this work, I present an end-to-end deep learning pipeline
+to estimate CO 2emissions, the most dominant greenhouse
+gas in the atmosphere, at an individual power-plant scale.
+My pipeline processes a single multi-spectral satellite image
+and associated weather data to extract smoke plumes from
+power plants and estimate power generation and CO 2emis-
+sion rates. The results of this work serve as a step towards the
+detection and monitoring of major sources of power plant
+GHG emissions on a global scale at a low-cost.
+Previous Works
+Previous works have explored the relations between plume
+imagery and GHG emission rates, and the applications of
+machine learning in predicting power plant behavior. Cus-
+worth et al. (Cusworth et al. 2021) employed airborne vis-
+ible/infrared imaging spectrometers (A VIRIS-NG) to quan-
+tify the carbon dioxide (CO 2) and methane (CH 4) emissions
+of 17 power plants from their smoke and vapor plumes.
+Aided by plant-specific annotations, Climate TRACE, a
+coalition working towards tracking all greenhouse gas emis-
+sions from anthropogenic activities, has been able to es-
+timate plant generation and emission rates from satellite
+imagery. Couture et al. (Couture et al. 2020) details Cli-
+mate TRACE’s methods in annotating cooling towers, flue
+stacks, and water outlets to aid in their model’s predic-
+tions. Both Cusworth and Climate TRACE’s respective ap-
+proaches are reliant on extensive data preparation and an-
+notation, thus making it difficult to produce a generaliz-
+able solution that can scale across large regions. More re-
+cently, Hannna et al. (Hanna et al. 2021) demonstrated the
+promise of using plume segmentation to inform more gen-
+eralizable model predictions, feeding a satellite image as an
+input to a pipeline capable of plume segmentation, power
+plant classification, and power generation prediction. This
+work builds off Hannna’s research by adding CO 2flux rates
+to the dataset, and comparing various state-of-the-art ma-
+chine learning architectures to produce a pipeline that per-
+forms well across the plume segmentation, fossil fuel clas-
+sification, power generation regression, and CO 2regression
+tasks.
+Methods
+Dataset
+The dataset from Hanna et al. is comprised of 2131 samples
+of multi-spectral satellite images taken by ESA’s Sentinel-
+2 satellites (Drusch et al. 2012). The resolution of the im-
+agery is 120px ×120px at 10m/px to cover an area of 1.2km
+×1.2km on the ground. Each image has a corresponding
+smoke plume mask that is used to train the segmentation
+models. These samples are paired with the plant’s longi-
+tude and latitude coordinates, country, weather data (tem-
+perature, humidity, wind), type of fossil fuel, and power
+generation rates. Using reported annual CO 2emissions and
+power plant generation capacities sourced from the Euro-
+pean Union Emissions Trading System (Verena Graichen
+2019), I convert the provided power generation rates into
+CO2emission, or flux, rates. The CO 2flux rate of the plantsranges from 307 tons/hour to 2834 tons/hour, with the aver-
+age flux rate being 1548 tons/hour.
+Data Preprocessing
+The data was split with 70% (1507 samples) going into the
+training set and 30% (624 samples) going in the testing sets
+such that each set did not contain images of the same power
+plant. All images in the dataset were normalized to reduce
+the effect of background objects or noise in the image. Then,
+all the images from the training set were duplicated five fold
+to increase the size of the training data to 7535 samples,
+and they all underwent a data augmentation process where
+they were randomly mirrored, flipped, cropped, and rotated
+a random amount between 0◦and 360◦both clockwise and
+counter-clockwise. This augmentation serves to generate a
+diverse set of possible plume orientations and center loca-
+tions that should help the model better generalize and pre-
+vent over fitting to the training set.
+Figure 1: Diagram of the model pipeline. It takes a multi-
+spectral satellite image as input and learns to do four tasks:
+(1) semantic segmentation of smoke plumes, (2) classifica-
+tion of type of fossil fuel, and (3) regression with respect to
+power generation and (4) CO 2emission rates.
+Model Pipeline
+The pipeline needs to accomplish four tasks: (1) semantic
+segmentation of smoke plumes in the satellite imagery, (2)
+classification of the type of fossil fuels being used by the
+power plant, (3) prediction of the plant’s power generation
+rate, and (4) prediction of the CO 2flux rate. Figure 1 shows
+the structure and flow of the model pipeline, and how the
+models for tasks 2-4 use outputs of other models to help
+inform their predictions. This is most significantly utilized
+for task 4, the prediction of the CO 2flux rate, which uses the
+output of all three previous tasks as input to the model. For
+
+each task, I evaluated 3 state-of-the-art model architectures
+that have shown to generally perform well in their respective
+tasks.
+Segmentation
+For the segmentation task, I chose FCN (Fully Convolu-
+tional Network), U-Net, and DeepLabV3 for experimenta-
+tion (Long, Shelhamer, and Darrell 2014), (Ronneberger,
+Fischer, and Brox 2015), (Chen et al. 2017). The FCN model
+consists of a set of max-pooling and convolution layers to
+identify and segment features in an image. The U-Net is
+based on FCN, but it employs an Encoder-Decoder architec-
+ture consisting of contracting and expanding convolutional
+layers. DeepLabv3 is a pre-trained model that also employs
+an encoder-decoder architecture with spatial pyramind pool-
+ing layers and atrous convolution techniques to learn about
+the larger context of the image it is segmenting. I mea-
+sure performance on this task using Intersection Over Union
+(IoU) and the loss function is binary cross entropy.
+Classification
+For classification, I employed transfer learning, and tested
+pre-trained models Res-Net 50, VGG-16, and InceptionV3,
+which were all created with different metrics to optimize
+(He et al. 2016), (Simonyan and Zisserman 2015), (Szegedy
+et al. 2016). ResNet prioritizes finding the simplest solu-
+tion through shortcut connections. VGG-16 is an optimized
+convolutional neural network model (CNN) with a focus on
+faster learning without over-fitting. InceptionV3 uses multi-
+ple kernal sizes to adapt to finding both larger, global fea-
+tures and smaller, area-specific features in an image, which
+is necessary for this task, as plumes can span across the en-
+tire satellite image or be a single spot in its corner. The cho-
+sen loss function is cross entropy loss, and the evaluation
+metric for this task is accuracy and Area Under the Curve
+(AUC).
+Regression
+Tasks 3 and 4 are regression problems, in which I evalu-
+ated Linear Regression, Artificial Neural Networks (ANN),
+and XGBoost (eXtreme Gradient Boost) (Chen and Guestrin
+2016). Linear Regression models the relationship between a
+set of variables through a linear equation. ANNs employ the
+neural network architecture and have done well in regression
+tasks. XGBoost is an implementation of the gradient boosted
+trees algorithm that learns to fit data by minimizing a regu-
+larized (L1 and L2) objective function. L1 loss was selected
+as the loss function and performance was measured through
+the R2coefficient, Mean Absolute Error (MAE) and Mean
+Absolute Percentage Error (MAPE).
+Results
+To train each model, I performed a hyperparameter search
+using the library Optuna, a framework that automates the
+training process by automatically adjusting the hyperparam-
+eters to maximize each of the listed performance metrics
+above (Akiba et al. 2019). The results from the training and
+test sets of all the models discussed is shown in Table 1.Table 1: Model Training Results
+Model Task Metric Train Test
+FCN Seg. IoU .752 .684
+DeepLabv3+ Seg. IoU .836 .769
+U-Net Seg. IoU .903 .841
+VGG-16 Cls. Acc. 76% 69%
+Inceptionv3 Cls. Acc. 86% 81%
+ResNet50 Cls. Acc. 94% 92%
+Lin Reg Pwr Reg. R2.803 .651
+ANN Pwr Reg. R2.837 .809
+XGBoost Pwr Reg. R2.893 .852
+Lin Reg Flux Reg. R2.723 .542
+ANN Flux Reg. R2.815 .748
+XGBoost Flux Reg. R2.861 .824
+Plume Segmentation
+The best performing segmentation model was the U-Net,
+achieving an IoU score of 0.903 on the training set and 0.841
+on the test set. The model performed very well on sam-
+ples where the plume masked the majority of the image, and
+performance declined on images with smaller plumes with
+more complicated shapes. I found that the model heavily uti-
+lized both associated weather data and certain multi-spectral
+imagery bands as key features that influenced its predictions.
+Particularly, the model used outside factors such as humid-
+ity and wind speeds to help it gain a larger context of the
+plume, and how it could have possibly been influenced by
+conditions that could not be captured by the satellite im-
+agery. Moreover, the Short-wave Infrared (SWIR) and Wa-
+ter Vapor imagery bands were able to capture thermal and
+visual details about the smoke plume that helped the model
+differentiate the plume from other background noise in the
+image, such as clouds, light buildings, or other terrain.
+Figure 2: Confusion Matrix of ResNet-50 Model for Fossil-
+Fuel Classification Task.
+
+Fossil Fuel Classification
+For fossil-fuel classification, the ResNet50 model reached
+an accuracy rate of 94% on the training and 92% on the test
+set, much higher than InceptionV3 and VGG-16. This model
+was able to generalize very well across the four classes, coal,
+peat, gas, and lignite, and the test set results are displayed
+in the Confusion Matrix (Figure 2). One possible source of
+bias in this model comes from the unequal distributions of
+classes in the dataset, where coal is present more than twice
+as much as peat.
+Power Plant Regression
+The XGBoost model outperformed Linear Regression and
+ANN, gaining a R2, MAPE of .861, 8.7% and .824, 10.2%
+on the training and test sets respectively. The output of the
+second model, the fossil-fuel classification prediction, had
+the most influence over these power generation predictions,
+as the per-hour CO 2emissions from coal power plants are
+much larger than the emissions from peat or natural gas
+power plants (Raghuvanshi, Chandra, and Raghav 2006).
+Initially, the model was largely over-fitting to the training
+set, and this was reduced through increased data augmenta-
+tion and the addition of several dropout layers, which both
+served as regularization techniques increasing the model’s
+variances to the training data. This enabled a better general
+fit and increased performance on the test set, where it was
+giving predictions on plants it had never seen before.
+CO 2Flux Rate Regression
+XGBoost was also the best performing model for CO 2flux
+rate regression, achieving an R2value of .824 and a MAPE
+of 10.8% on the test set. Figure 3 exhibits this high perfor-
+mance, where the .87 line slope indicates a high correlation
+between the model’s predictions and the ground truth data.
+Model performance on the CO 2emission rate predictions
+was heavily dependent on the accuracy of the power genera-
+tion predictions, as seen from the direct relationship between
+power generation and CO 2flux rate mentioned above. The
+XGBoost model was able to generalize very well to the data,
+and it is a promising algorithm to further evaluate to see if it
+can continue to perform well across other regions.
+Conclusions and Future Work
+In this work, I developed an end-to-end deep learning
+pipeline that successfully predicted power generation and
+CO2emission rates across Europe via high resolution re-
+mote sensing data, an important step toward a future of ac-
+curate emissions monitoring across the globe. My pipeline
+performed well across all of its tasks (plume segmentation,
+fossil-fuel classification, power generation regression, and
+CO2flux rate regression) and demonstrates the promise of
+the plume segmentation approach acting as a possible gen-
+eralizable solution to measure emissions across many power
+plants.
+This project identified a number of features, techniques,
+and models that hold promise for evaluation in future works.
+The use of Shortwave Infrared (SWIR) imagery for differen-
+tiating plumes and other pollutants from background noise
+Figure 3: XGBoost Model Predicted CO 2Emissions v.s.
+Ground Truth CO 2Emissions (Flux Rate).
+can serve as a key component for creating adaptive models
+to generalize to regional patterns and operate at night. The
+application of XGBoost in regression tasks can be further
+evaluated to see if the model can maintain its high accuracy
+rates across a larger sample size of data. Data accessibil-
+ity remains a key component to the expansion of this work.
+Currently, the model has only trained on European power
+plants, and additional testing is required to measure model
+bias and see if this performance can translate to other re-
+gions and countries, such as the United States and China, in
+order for it to be truly globally scalable. In the near future, I
+aim to make this work more accurate and precise, with a fo-
+cus on expanding to lesser-developed regions such as India
+and Brazil. Moreover, as more data becomes available, the
+pipeline can extended to predict other gases, such as such as
+methane (CH 4) and nitrous oxide (N 2O).
+Global emissions monitoring systems will radicalize cli-
+mate action efforts, providing a new level of reliable and
+transparent data that can aid governments and companies in
+designing effective climate policy. For example, by helping
+to identify “super-emitter” power plants, this pipeline pin-
+points locations where government regulation is necessary
+and renewable alternatives will have the most impact. The
+results of this work serve as a step toward the low-cost mon-
+itoring and detection of major sources of GHG emissions,
+helping limit their catastrophic impacts on climate and our
+planet.
+Acknowledgments
+Part of this research was done in affiliation with WattTime,
+a member of the Climate TRACE coalition. I would like to
+thank Hannes Koenig, Jeremy Freeman, Heather Couture,
+and everyone else at WattTime for their help and mentorship
+that aided in the development of this work.
+
+References
+Akiba, T.; Sano, S.; Yanase, T.; Ohta, T.; and Koyama, M.
+2019. Optuna: A Next-generation Hyperparameter Opti-
+mization Framework. arXiv:1907.10902 [cs, stat] . ArXiv:
+1907.10902.
+Apte, J. S.; Messier, K. P.; Gani, S.; Brauer, M.; Kirchstetter,
+T. W.; Lunden, M. M.; Marshall, J. D.; Portier, C. J.; Ver-
+meulen, R. C.; and Hamburg, S. P. 2017. High-Resolution
+Air Pollution Mapping with Google Street View Cars: Ex-
+ploiting Big Data. Environmental Science & Technology ,
+51(12): 6999–7008. Publisher: American Chemical Society.
+Boesch, H.; Liu, Y .; Tamminen, J.; Yang, D.; Palmer, P. I.;
+Lindqvist, H.; Cai, Z.; Che, K.; Di Noia, A.; Feng, L.;
+Hakkarainen, J.; Ialongo, I.; Kalaitzi, N.; Karppinen, T.;
+Kivi, R.; Kivim ¨aki, E.; Parker, R. J.; Preval, S.; Wang, J.;
+Webb, A. J.; Yao, L.; and Chen, H. 2021. Monitoring Green-
+house Gases from Space. Remote Sensing , 13(14): 2700.
+Number: 14 Publisher: Multidisciplinary Digital Publishing
+Institute.
+Chen, L.-C.; Papandreou, G.; Schroff, F.; and Adam, H.
+2017. Rethinking Atrous Convolution for Semantic Image
+Segmentation.
+Chen, T.; and Guestrin, C. 2016. XGBoost: A Scalable
+Tree Boosting System. In Proceedings of the 22nd ACM
+SIGKDD International Conference on Knowledge Discov-
+ery and Data Mining , KDD ’16, 785–794. New York, NY ,
+USA: Association for Computing Machinery. ISBN 978-1-
+4503-4232-2.
+Couture, H. D.; O’Connor, J.; Mitchell, G.; S ¨oldner-
+Rembold, I.; D’souza, D.; Karra, K.; Zhang, K.;
+Rouzbeh Kargar, A.; Kassel, T.; Goldman, B.; Tyrrell,
+D.; Czerwinski, W.; Talekar, A.; and McCormick, C. 2020.
+Towards Tracking the Emissions of Every Power Plant on
+the Planet. In NeurIPS 2020 Workshop on Tackling Climate
+Change with Machine Learning .
+Cusworth, D. H.; Duren, R. M.; Thorpe, A. K.; East-
+wood, M. L.; Green, R. O.; Dennison, P. E.; Franken-
+berg, C.; Heckler, J. W.; Asner, G. P.; and Miller,
+C. E. 2021. Quantifying Global Power Plant Car-
+bon Dioxide Emissions With Imaging Spectroscopy.
+AGU Advances , 2(2): e2020A V000350. eprint:
+https://onlinelibrary.wiley.com/doi/pdf/10.1029/2020A V000350.
+Drusch, M.; Del Bello, U.; Carlier, S.; Colin, O.; Fernan-
+dez, V .; Gascon, F.; Hoersch, B.; Isola, C.; Laberinti, P.;
+Martimort, P.; Meygret, A.; Spoto, F.; Sy, O.; Marchese, F.;
+and Bargellini, P. 2012. Sentinel-2: ESA’s Optical High-
+Resolution Mission for GMES Operational Services. Re-
+mote Sensing of Environment , 120: 25–36.
+Eldering, A.; Wennberg, P.; Crisp, D.; Schimel, D.; Gun-
+son, M.; Chatterjee, A.; Liu, J.; Schwandner, F. M.; Sun,
+Y .; O’Dell, C.; Frankenberg, C.; Taylor, T.; Fisher, B.; Os-
+terman, G.; Wunch, D.; Hakkarainen, J.; Tamminen, J.; and
+Weir, B. 2017. The Orbiting Carbon Observatory-2 early
+science investigations of regional carbon dioxide fluxes. Sci-
+ence (New York, N.Y.) , 358(6360): eaam5745.
+Hanna, J.; Mommert, M.; Scheibenreif, L. M.; and Borth, D.
+2021. Multitask Learning for Estimating Power Plant Green-house Gas Emissions from Satellite Imagery. In NeurIPS
+2021 Workshop on Tackling Climate Change with Machine
+Learning .
+He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep Residual
+Learning for Image Recognition. In 2016 IEEE Conference
+on Computer Vision and Pattern Recognition (CVPR) , 770–
+778. Las Vegas, NV , USA: IEEE. ISBN 978-1-4673-8851-1.
+Long, J.; Shelhamer, E.; and Darrell, T. 2014. Fully Convo-
+lutional Networks for Semantic Segmentation.
+Mooney, C.; Eilperin, J.; Butler, D.; Muyskens, J.; Narayan-
+swamy, A.; and Ahmed, N. ???? Countries’ climate pledges
+built on flawed data, Post investigation finds.
+on Climate Change, I. P.; and Edenhofer, O., eds. 2014. Cli-
+mate change 2014: mitigation of climate change: Working
+Group III contribution to the Fifth Assessment Report of the
+Intergovernmental Panel on Climate Change . New York,
+NY: Cambridge University Press. ISBN 978-1-107-05821-7
+978-1-107-65481-5. OCLC: ocn892580682.
+Raghuvanshi, S. P.; Chandra, A.; and Raghav, A. K. 2006.
+Carbon dioxide emissions from coal based power generation
+in India. Energy Conversion and Management , 47(4): 427–
+441.
+Ronneberger, O.; Fischer, P.; and Brox, T. 2015. U-Net:
+Convolutional Networks for Biomedical Image Segmenta-
+tion. arXiv:1505.04597 [cs] . ArXiv: 1505.04597.
+Simonyan, K.; and Zisserman, A. 2015. Very Deep Con-
+volutional Networks for Large-Scale Image Recognition.
+ArXiv:1409.1556 [cs].
+Smol, J. P. 2012. Climate Change: A planet in flux. Na-
+ture, 483(7387): S12–S15. Number: 7387 Publisher: Nature
+Publishing Group.
+Szegedy, C.; Vanhoucke, V .; Ioffe, S.; Shlens, J.; and Wojna,
+Z. 2016. Rethinking the Inception Architecture for Com-
+puter Vision. In 2016 IEEE Conference on Computer Vi-
+sion and Pattern Recognition (CVPR) , 2818–2826. Las Ve-
+gas, NV , USA: IEEE. ISBN 978-1-4673-8851-1.
+US EPA, O. 2016. EMC: Continuous Emission Monitoring
+Systems.
+Verena Graichen, S. G., Johanna Cludius. 2019. Euro-
+pean Union Emissions Trading System (EU ETS) data from
+EUTL — European Environment Agency.
+

aaaifss2022_12.txt

@@ -0,0 +1,928 @@
+CLIMATE BERT: A Pretrained Language Model for Climate-Related Text
+Nicolas Webersinke,1Mathias Kraus,1Julia Anna Bingler,2Markus Leippold3
+1FAU Erlangen-Nuremberg, Germany
+2ETH Zurich, Switzerland
+3University of Zurich, Switzerland
+[email protected], [email protected], [email protected], [email protected]
+Abstract
+Over the recent years, large pretrained language models (LM)
+have revolutionized the field of natural language processing
+(NLP). However, while pretraining on general language has
+been shown to work very well for common language, it has
+been observed that niche language poses problems. In par-
+ticular, climate-related texts include specific language that
+common LMs can not represent accurately. We argue that
+this shortcoming of today’s LMs limits the applicability of
+modern NLP to the broad field of text processing of climate-
+related texts. As a remedy, we propose C LIMATE BERT, a
+transformer-based language model that is further pretrained
+on over 2 million paragraphs of climate-related texts, crawled
+from various sources such as common news, research arti-
+cles, and climate reporting of companies. We find that C LI-
+MATE BERT leads to a 48% improvement on a masked lan-
+guage model objective which, in turn, leads to lowering error
+rates by 3.57% to 35.71% for various climate-related down-
+stream tasks like text classification, sentiment analysis, and
+fact-checking.
+1 Introduction
+Researchers working on climate change-related topics in-
+creasingly use natural language processing (NLP) to auto-
+matically extract relevant information from textual data. Ex-
+amples include the sentiment or specificity of language used
+by companies when discussing climate risks and measuring
+corporate climate change exposure, which increases trans-
+parency to help the public know where we stand on climate
+change (e.g., Callaghan et al. 2021; Bingler et al. 2022b).
+Many studies in this domain apply traditional NLP meth-
+ods, such as dictionaries, bag-of-words approaches or sim-
+ple extensions thereof (e.g., Gr ¨uning 2011; Sautner et al.
+2022). However, such analyses face considerable limita-
+tions, since climate-related wording could vary substan-
+tially by source (Kim and Kang 2018). Deep learning tech-
+niques that promise higher accuracy are gradually replacing
+these approaches (e.g., K ¨olbel et al. 2020; Luccioni, Baylor,
+and Duchene 2020; Bingler et al. 2022a; Callaghan et al.
+2021; Wang, Chillrud, and McKeown 2021; Friederich et al.
+2021). Indeed, it has been shown in related domains that
+Copyright c
+Intelligence (www.aaai.org). All rights reserved.deep learning in NLP allows for impressive results, outper-
+forming traditional methods by large margins (Varini et al.
+2020).
+These deep learning-based approaches make use of lan-
+guage models (LMs), which are trained on large amounts
+of textual and unlabelled data. This training on unlabelled
+data is called pretraining and leads to the model learning
+representations of words and patterns of common language.
+One of the most prominent language models is called BERT
+(Bidirectional Encoder Representations from Transformers)
+(Devlin et al. 2018) with its successors R OBERT A(Liu et al.
+2019), Transformer-XL (Dai et al. 2019) and ELECTRA
+(Clark et al. 2020). These models have been trained on huge
+amounts of text which was crawled from an unprecedented
+amount of online resources.
+After the pretraining phase, most LMs are trained on addi-
+tional tasks, the downstream task . For the downstream tasks,
+the LM builds on and benefits from the word representations
+and language patterns learned in the pretraining phase. The
+pre-training benefit is especially large on downstream tasks
+for which the collection of samples is difficult and, thus, the
+resulting training datasets are small (hundreds or few thou-
+sands of samples). Furthermore, it has been shown that a
+model that was pretrained on the downstream task-specific
+text exhibits better performance, compared to a model that
+has been pretrained solely on general text (Araci 2019; Lee
+et al. 2020).
+Hence, a straightforward extension to the standard com-
+bination of pretraining is the so-called domain-adaptive pre-
+training (Gururangan et al. 2020). This approach has re-
+cently been studied for various tasks and basically comes in
+the form of pretraining multiple times — in particular pre-
+training in the language domain of the downstream task, i.e.,
+pretraining (general domain)
++domain-adaptive
+pretraining (downstream domain)
++training (downstream task) :
+To date, regardless of the increase in using NLP for cli-
+mate change related research, a model with climate domain-
+adaptive pretraining has not been publicly available, yet.
+Research so far rather relied on models pretrained on gen-
+eral language, and fine-tuned on the downstream task. To
+
+fill this gap, our contribution is threefold. First, we in-
+troduce C LIMATE BERT, a state-of-the-art language model
+that is specifically pretrained on climate-related text cor-
+pora of various sources, namely news, corporate disclosures,
+and scientific articles. This language model is designed to
+support researchers of various disciplines in obtaining bet-
+ter performing NLP models for a manifold of downstream
+tasks in the climate change domain. Second, to illustrate
+the strength of C LIMATE BERT, we highlight the perfor-
+mance improvements using C LIMATE BERT on three stan-
+dard climate-related NLP downstream tasks. Third, to fur-
+ther promote research at the intersection of climate change
+and NLP, we make the training code and weights of all lan-
+guage models publicly available at GitHub and Hugging
+Face.12
+2 Background
+As illustrated in Figure 1, our LM training approach for C LI-
+MATE BERTcomprises all three phases — using an LM pre-
+trained on a general domain, the domain-adaptive pretrain-
+ing on the climate domain, and the training phase on climate-
+related downstream tasks.
+Pretraining on General Domain
+As of 2018, pretraining became the quasi-standard for learn-
+ing NLP models. First, a neural language model, often with
+millions of parameters, is trained on large unlabeled corpora
+in a semi-supervised fashion. By learning on multiple levels
+which words/word-sequences/sentences appear in the same
+context, an LM can represent a semantically similar text by
+similar vectors. Typical objectives for training LMs are the
+prediction of masked words or the prediction of a label indi-
+cating whether two sentences occurred consecutively in the
+corpora (Devlin et al. 2018).
+In the earlier NLP pretraining days, LMs tradition-
+ally used regular or convolutional neural networks (Col-
+lobert and Weston 2008), or later Long-Short-Term-Memory
+(LSTM) networks to process text (Howard and Ruder 2018).
+Todays LMs mostly build on transformer models (e.g., De-
+vlin et al. 2018; Dai et al. 2019; Liu et al. 2019). One of
+the latter is named R OBERT A(Liu et al. 2019) which was
+trained on 160GB of various English-language corpora -
+data from BOOKCORPUS (Zhu et al. 2015), WIKIPEDIA,
+a portion of the CCNEWS dataset (Nagel 2016), OPEN-
+WEBTEXT corpus of web content extracted from URLs
+shared on Reddit (Gokaslan and Cohen 2019), and a sub-
+set of CommonCrawl that is said to resemble the story-like
+style of WINOGRAD schemas (Trinh and Le 2019). While
+these sources are valuable to build a model working on gen-
+eral language, it has been shown that domain-specific, niche
+language (such as climate-related text) poses a problem to
+current state-of-the-art language models (Araci 2019).
+Domain-Specific Pretraining
+As a remedy to inferior performance of general language
+models when applied to niche topics, multiple language
+1www.github.com/climatebert/language-model
+2www.huggingface.co/climatebertmodels have been proposed which build on the pretrained
+models but continue pretraining on their respective domains.
+FinBERT, LegalBert, MedBert are just a few language mod-
+els that have been further pretrained on the finance, legal, or
+medical domain (Araci 2019; Chalkidis et al. 2020; Rasmy
+et al. 2021). In general, this domain-adaptive pretraining
+yields more accurate models on downstream tasks (Guru-
+rangan et al. 2020).
+Domain-specific pretraining requires a decision about
+which samples to include in the training process. It is still an
+open debate which sample strategy improves performance
+best. Various strategies can be applied to extract the text
+samples on which the LM is further pretrained. For exam-
+ple, while traditional pretraining uses all samples from the
+pretraining corpus, similar sample selection (S IM-SELECT )
+uses only a subset of the corpus, in which the samples are
+similar to the samples in the downstream task (Ruder and
+Plank 2017). In contrast, diverse sample selection (D IV-
+SELECT ) uses a subset of the corpus, which includes dissim-
+ilar samples compared to the downstream dataset (Ruder and
+Plank 2017). Previous research has investigated the benefit
+of these approaches, yet no final conclusion about the effi-
+ciency has been obtained. Consequently, we compare these
+approaches in our experiments.
+NLP on Climate-Related Text
+In the past, climate-related textual analysis often used pre-
+defined dictionaries of presumably relevant words and then
+simply searched for these words within the documents.
+For example, Cody et al. (2015) use such an approach
+for climate-related tweets. Similarly, Sautner et al. (2022)
+use a keyword-based approach to capture firm-level climate
+change exposure. However, these methods do not account
+for context. The lack of context is a significant drawback,
+given the ambiguity of many climate-related words such
+as ”environment,” ”sustainable,” or ”climate” itself (Varini
+et al. 2020).
+Only recently, BERT has been used for NLP in climate-
+related text. The transformers-based BERT models are ca-
+pable of accounting for the context of words and have out-
+performed traditional approaches by large margins across
+various climate-related datasets (K ¨olbel et al. 2020; Luc-
+cioni, Baylor, and Duchene 2020; Varini et al. 2020; Bin-
+gler et al. 2022a; Callaghan et al. 2021; Wang, Chillrud, and
+McKeown 2021; Friederich et al. 2021; Stammbach et al.
+2022). However, this research has also shown that extracting
+climate-related information from textual sources is a chal-
+lenge, as climate change is a complex, fast-moving, and of-
+ten ambiguous topic with scarce resources for popular text-
+based AI tasks.
+While context-based algorithms like BERT can detect
+a variety of complex and implicit topic patterns in addi-
+tion to many trivial cases, there remains great potential
+for improvement in several directions. To our knowledge,
+none of the above cited work has examined the effects of
+domain-adaptive pretraining on their specific downstream
+tasks. Therefore, we investigate whether domain-adaptive
+pretraining will improve performance for climate change-
+related downstream tasks such as text classification, senti-
+
+News Abstracts ReportsCommon
+crawlPretraining (general domain)Domain-adaptive pretraining (climate
+domain)Training (downstream tasks)
++ +- Text classification
+- Sentiment analysis
+- Fact-checkingFigure 1: Sequence of training phases. Our main contribution is the continued pretraining of language models on the climate
+domain. In addition, we evaluate the obtained climate domain-specific language models on various downstream tasks.
+ment analysis, and fact-checking.
+3 C LIMATE BERT
+In the following, we describe our approach to train C LI-
+MATE BERT. We first list the underlying data sources before
+describing our sample selection techniques and, finally, the
+vocabulary augmentation we used for training the language
+model.
+Text Corpus
+Our goal was to collect a large corpus of text, C ORP, that
+included general and domain-specific climate-related lan-
+guage. We decided to include the following three sources:
+news articles, research abstracts, and corporate climate re-
+ports. We decided not to include full research articles be-
+cause this language is likely too specific and does not rep-
+resent general climate language. We also did not include
+Twitter data, as we assume that these texts are too noisy. In
+total, we collected 2,046,523 paragraphs of climate-related
+text (see Table 1).
+The N EWS dataset is mainly retrieved from Refinitiv
+Workspace and includes 135,391 articles tagged with cli-
+mate change topics such as climate politics, climate actions,
+and floods and droughts. In addition, we crawled climate-
+related news articles from the web.
+The A BSTRACTS dataset includes abstracts of climate-
+related research articles crawled from the Web of Science,
+primarily published between 2000 and 2019.
+The R EPORTS dataset comprises corporate climate and
+sustainability reports of more than 600 companies from the
+years 2015-2020 retrieved from Refinitiv Workspace and the
+respective company websites.
+Given the nature of the datasets, we find a large het-
+erogeneity between the paragraphs in terms of number
+of words. Unsurprisingly, on average, the paragraphs with
+the least words come from the N EWS and the R EPORTS
+datasets. In contrast, A BSTRACTS includes paragraphs with
+the most words. Table 1 lists these descriptives.
+To estimate the benefit from domain-adaptive pretrain-
+ing, we compare the similarity of our text corpus with the
+one used for pretraining R OBERT A. Following Gururangan
+et al. (2020), we consider the vocabulary overlap between
+both corpora. The resulting overlap of 57.05% highlights the
+dissimilarity between the two domains and the need to add
+specific vocabularies. Therefore, we expect to see consid-
+erable performance improvements of domain-adaptive pre-
+training.Dataset Num. of Avg. num. of words
+paragraphs Q1 Mean Q3
+News 1,025,412 34 56 65
+Abstracts 530,819 165 218 260
+Reports 490,292 34 65 79
+Total 2,046,523 36 107 168
+Table 1: Corpus C ORP used for pretraining C LIMATE BERT.
+Q1 and Q3 stand for the 0.25 and 0.75 quantiles, respec-
+tively.
+Sample Selection
+Prior work has shown that specific selections of the samples
+used for pretraining can foster the performance of the LM. In
+particular, incorporating information from the downstream
+task by selecting similar or diverse samples has been shown
+to yield favorable results compared to using all samples from
+the dataset. We follow both approaches and select samples
+that are similar or diverse to climate-text using our text clas-
+sification task (see 5). We experiment with three different
+strategies from Ruder and Plank (2017) for the selection of
+samples from our corpus:
+In the most traditional sample selection strategy, F ULL-
+SELECT , we use all paragraphs from C ORP to train
+CLIMATE BERTF.
+In S IM-SELECT , we select the 70% of samples from
+CORP, which are most similar to the samples of our text
+classification task. We use a Euclidean similarity met-
+ric for this sample selection strategy. We call this LM
+CLIMATE BERTS.
+In D IV-SELECT , we select the 70% of samples from
+CORP, which are most diverse compared to the samples
+from our text classification task. We use the sum be-
+tween the type-token-ratio and the Shannon-entropy for
+measuring diversity (Ruder and Plank 2017). This LM is
+named C LIMATE BERTD.
+In D IV-SELECT + SIM-SELECT , we use the same diver-
+sity and similarity metrics as before. We then compute
+a composite score by summing over their scaled values.
+We keep the 70% of the samples with the highest com-
+posite score to train C LIMATE BERTD+S.
+
+Downstream domain- Downstream tasks
+adaptive pretraining training
+Hyperparameter Value
+Batch size 2016 32
+Learning rate 5e-4 5e-5
+Number of epochs 12 1000
+Patience — 4
+Class weight — Balanced
+Feedforward nonlinearity — tanh
+Feedforward layer — 1
+Output neurons — Task dependent
+Optimizer Adam
+Adam epsilon 1e-6
+Adam beta weights (0.9, 0.999)
+Learning rate scheduler Warmup linear
+Weight decay 0.01
+Table 2: Hyperparameters used for the downstream domain-
+adaptive pretraining and the downstream tasks training of
+CLIMATE BERT.
+Vocabulary Augmentation
+We extend the existing vocabulary of the original model
+to include domain-specific terminology. This allows C LI-
+MATE BERT to explicitly learn representations of terminol-
+ogy that frequently occur in a climate-related text but not in
+the general domain. In particular, we add the 235 most com-
+mon tokens as new tokens to the tokenizer, thereby extend-
+ing the size of the vocabulary for our basis language model
+(DistilR OBERT A) from 50,265 to 50,500. See Appendix C
+for a list of all added tokens. We also experimented with
+language models that do not use vocabulary augmentation
+or add more tokens. However, overall we find improvements
+using this technique and, thus, apply it to all language mod-
+els which we pretrain on the climate domain.
+Model Selection
+For all our experiments, we use DistilR OBERT A, a distilled
+version of R OBERT Afrom Huggingface,3as our starting
+point for training (Sanh et al. 2019). All our language mod-
+els are trained with a masked language modeling objective
+(i.e., cross-entropy loss on predicting randomly masked to-
+kens). We report all hyperparameters in Table 2. The large
+batch size of 2016 for training the LM is achieved using gra-
+dient accumulation.
+Training on Downstream Task
+After pretraining DistilR OBERT Aon C ORP, we follow
+standard practice (Devlin et al. 2018) and pass the final layer
+[CLS] token representation to a task-specific feedforward
+layer for prediction. We report all hyperparameters of this
+feedforward layer in Table 2.
+4 Performance Analysis of Language Model
+Table 3 lists the results after pretraining DistilR OBERT Aon
+CORP with various sample selection strategies. For evalu-
+ation, we split C ORP randomly into 80% training data and
+20% validation data. The reported loss is the cross-entropy
+3www.huggingface.co/distilroberta-baseloss on predicting randomly masked tokens from the valida-
+tion data. We find that C LIMATE BERTFleads to the lowest
+validation loss. This performance is followed by the other
+CLIMATE BERT LMs, which all show similar results. Over-
+all, we find that our domain-adaptive pretraining decreases
+the cross-entropy loss by 46–48% compared to the basis Dis-
+tilR OBERT A, cutting the loss almost in half.
+Model Val. loss
+DistilR OBERT A 2.238
+CLIMATE BERTF 1.157
+CLIMATE BERTS 1.205
+CLIMATE BERTD 1.204
+CLIMATE BERTD+S 1.203
+Table 3: Loss of our language models on a validation set
+from our text corpus C ORP.
+5 Performance Analysis for Climate-Related
+Downstream Tasks
+For our experiments, we used the following downstream
+tasks: text classification, sentiment analysis, and fact-
+checking. Table 4 lists basic statistics about the downstream
+tasks. We repeated the training and evaluation phase 60
+times for each experiment, each time with a random 90%
+set of samples for training and the remaining 10% set for
+validation.
+Downstream Num. of Labels Label
+task samples distribution
+Text classification 1220 climate-related: yes/no 1000/220
+Sentiment analysis 1000 opportunity/neutral/risk 250/408/342
+Fact-checking 2745 claim: support/refute 1943/802
+Table 4: Overview of our downstream tasks used for evalu-
+ating C LIMATE BERT.
+Text Classification
+For our text classification experiment, we use a dataset con-
+sisting of hand-selected paragraphs from companies’ annual
+reports or sustainability reports. All paragraphs were anno-
+tated as yes(climate-related) or no(not climate-related) by at
+least four experts from the field using the software prodigy.4
+See Appendix B for our annotation guidelines. In case of a
+close verdict or a tie between the annotators, the authors of
+this paper discussed the paragraph in depth before reaching
+an agreement.
+In the following, Table 5 reports the results of the lan-
+guage models when trained on our text classification task,
+i.e., whether the text is climate-related or not. Overall,
+we find that all C LIMATE BERT LMs outperform a non-
+pre-trained DistilR OBERT Aacross both metrics for the
+text classification task. Most notably, domain-adaptive pre-
+training with similar samples to our downstream tasks
+4www.prodi.gy
+
+(CLIMATE BERTS) leads to improvements of 32.64% in
+terms of cross-entropy loss and a reduction in the error rate
+of the F1 score by 35.71%.
+Text classification
+Model Loss F1
+DistilR OBERT A 0:242 0:171 0:986 0:010
+CLIMATE BERTF 0:191 0:136 0:989 0:010
+CLIMATE BERTS 0:163 0:132 0:991 0:008
+CLIMATE BERTD 0:197 0:153 0:988 0:009
+CLIMATE BERTD+S0:217 0:153 0:988 0:009
+Table 5: Results on our text classification task. Reported are
+the average cross-entropy loss and the average weighted F1
+score on the validation sets across 60 evaluation runs. Value
+subscripts report the standard deviations.
+Sentiment Analysis
+Our next task studies the sentiment behind the climate-
+related paragraphs, using the same dataset as in the previ-
+ous section. In our context, we use the term ‘sentiment’ to
+distinguish whether an entity reports on climate-related de-
+velopments as negative risk, as positive opportunity , or as
+neutral .
+Therefore, we created a second labeled dataset on climate-
+related sentiment, for which we asked the annotators to label
+the paragraphs by one of the three categories — risk,neutral ,
+oropportunity . See Appendix B for our annotation guide-
+lines. Similarly, as before, in case of a close verdict or a tie
+between the annotators, the authors of this paper discussed
+the paragraph in depth before reaching an agreement.
+Table 6 shows the performance of our models in senti-
+ment prediction. Again, all C LIMATE BERTLMs outperform
+the DistilR OBERT Abaseline model in terms of F1 score and
+average cross-entropy loss. The largest improvements can be
+observed with C LIMATE BERTF, which amount to a 7.33%
+lower cross-entropy loss and a 7.42% lower error rate in
+terms of average F1 score compared to the DistilR OBERT A
+baseline LM.
+Sentiment analysis
+Model Loss F1
+DistilR OBERT A 0:150 0:069 0:825 0:046
+CLIMATE BERTF 0:139 0:042 0:838 0:036
+CLIMATE BERTS 0:140 0:057 0:836 0:033
+CLIMATE BERTD 0:138 0:043 0:835 0:040
+CLIMATE BERTD+S0:139 0:043 0:834 0:036
+Table 6: Results on our sentiment analysis task in terms
+of average validation loss and average weighted F1 score
+across 60 evaluation runs. Subscripts report the standard de-
+viations.Fact-Checking
+We now turn to the fact-checking downstream task. We ap-
+ply our model to a dataset that was proposed by Diggelmann
+et al. (2020) and comprises 1.5k sentences that make a claim
+about climate-related topics. This CLIMATE -FEVER dataset
+is to the best of our knowledge to date the only dataset
+that focuses on climate change fact-checking. CLIMATE -
+FEVER adapts the methodology of FEVER , the largest dataset
+of artificially designed claims, to real-life claims on cli-
+mate change collected online. The authors of CLIMATE -
+FEVER find that the surprising, subtle complexity of mod-
+eling real-world climate-related claims provides a valuable
+challenge for general natural language understanding. Work-
+ing with this dataset, Wang, Chillrud, and McKeown (2021)
+recently introduced a novel semi-supervised training method
+to achieve a state-of-the-art (SotA) F1 score of 0.7182 on the
+fact-checking dataset CLIMATE -FEVER .
+Claim : 97% consensus on human-caused
+global warming has been disproven.
+Evidence
+REFUTE: In a 2019 CBS poll, 64% of the US
+population said that climate change
+is a ””crisis”” or a ””serious prob-
+lem””, with 44% saying human ac-
+tivity was a significant contributor.
+Claim : The melting Greenland ice sheet is
+already a major contributor to ris-
+ing sea level and if it was eventu-
+ally lost entirely, the oceans would
+rise by six metres around the world,
+flooding many of the world’s largest
+cities.
+Evidence
+SUPPORT: The Greenland ice sheet occupies
+about 82% of the surface of Green-
+land, and if melted would cause sea
+levels to rise by 7.2 metres.
+Table 7: Examples taken from CLIMATE -FEVER .
+Each claim in CLIMATE -FEVER is supported or refuted by
+evidence sentences (see Table 7), and an evidence sentence
+can also be classified as giving not enough information. The
+objective of the model is to classify an evidence sentence to
+support orrefute a claim. To feed this combination of claim
+and evidence into the model, we concatenate the claims with
+the related evidence sentences, with a [SEP] token sepa-
+rating them. As in Wang, Chillrud, and McKeown (2021),
+and for comparison with their results, we filter out all evi-
+dence sentences with the label NOT ENOUGH INFO in the
+CLIMATE -FEVER dataset.
+Table 8 lists the results of our experiments on the
+CLIMATE -FEVER dataset. In line with our previous exper-
+iments, we find similar or better results for all C LIMATE -
+BERTLMs across all metrics. Our C LIMATE BERTD+SLM
+achieves similar cross-entropy loss compared to the basis
+DistilR OBERT Amodel, yet pushes the average F1 score
+from 0.748 to 0.757, which outperforms Wang, Chillrud, and
+McKeown (2021)’s previous SotA F1 score of 0.7182, and
+
+is hence, to the best of our knowledge, the new SotA on this
+dataset.
+Fact-checking
+Model Loss F1
+DistilR OBERT A 0:135 0:017 0:748 0:036
+CLIMATE BERTF 0:134 0:020 0:755 0:037
+CLIMATE BERTS 0:133 0:017 0:753 0:042
+CLIMATE BERTD 0:135 0:016 0:752 0:042
+CLIMATE BERTD+S0:135 0:018 0:757 0:044
+Table 8: Results on our fact-checking task on CLIMATE -
+FEVER in terms of average validation loss and average
+weighted F1 score across 60 evaluation runs. Subscripts re-
+port the standard deviations.
+6 Carbon Footprint
+Training deep neural networks in general and large lan-
+guage models in particular, has a significant carbon footprint
+already today. If the LM research trends continue, this detri-
+mental climate impact will increase considerably. The topic
+of efficient NLP was also discussed by a working group
+appointed by the ACL Executive Committee to promote
+ways that the ACL community can reduce the computational
+costs of model training (https://public.ukp.informatik.tu-
+darmstadt.de/enlp/Efficient-NLP-policy-document.pdf).
+We acknowledge that our work is part of this trend. In
+total, training C LIMATE BERTcaused 115.15 kg CO2 emis-
+sions. We use two energy efficient NVIDIA RTX A5000
+GPUs: 0.7 kW (power consumption of GPU server) x 350
+hours (combined training time of all experiments) x 470
+gCO2e/kWh (emission factor in Germany in 2018 according
+to www.umweltbundesamt.de/publikationen/entwicklung-
+der-spezifischen-kohlendioxid-7) = 115,149 gCO2e. We
+list all details about our climate impact in Table 9 in
+Appendix A. Nevertheless, we decided to carry out this
+project, as we see the high potential of NLP to support
+action against climate change. Given our awareness of the
+carbon footprint of our research, we address this sensitive
+topic as follows:
+1. We specifically decided to focus on DistilR OBERT A,
+which is a considerably smaller model in terms of num-
+ber of parameters compared to the non-distilled version
+and, thus, requires less energy to train. Moreover, we do
+not crawl huge amounts of data without considering the
+quality. This way, we try to take into account the issues
+mentioned by Bender et al. (2021).
+2. Hyperparameter tuning yields considerably higher CO2
+emissions in the training stage due to tens or hundreds
+of different training runs. Note that our multiple train-
+ing runs on the downstream task are not causing long
+training times as the downstream datasets are very small
+compared to the dataset used for training the language
+model. We therefore refrain from exhaustive hyperpa-
+rameter tuning. Rather, we build on previous findings.We systematically experimented with a few hyperparam-
+eter combinations and found that the hyperparameters
+proposed by Gururangan et al. (2020) lead to the best
+results.
+3. We would have liked to train and run our model on
+servers powered by renewable energy. This first best op-
+tion was unfortunately not available. In order to speed
+up the energy system transformation required to achieve
+the global climate targets, we contribute our part by do-
+nating Euro 100 to atmosfair. atmosfair was founded in
+2005 and is supported by the German Federal Environ-
+ment Agency. atmosfair offsets carbon dioxide in more
+than 20 locations: from efficient cookstoves in Nigeria,
+Ethiopia and India to biogas plants in Nepal and Thai-
+land to solar energy in Senegal and Brazil and renewable
+energies in Tansania and Indonesia. See www.atmosfair.
+de/en/offset/fix/. We explicitly refrain from calling this
+donation a CO2 compensation, and we refrain from a so-
+lution that is based on afforestation.
+7 Conclusion
+We propose C LIMATE BERT, the first language model that
+was pretrained on a large scale dataset of over 2 mil-
+lion climate-related paragraphs. We study various selec-
+tion strategies to find samples from our corpus which are
+most helpful for later tasks. Our experiments reveal that
+our domain-adaptive pretraining leads to considerably lower
+masked language modeling loss on our climate corpus. We
+further find that this improvement is also reflected in predic-
+tive performance across three essential downstream climate-
+related NLP tasks: text classification, the analysis of risk and
+opportunity statements by corporations, and fact-checking
+climate-related claims.
+Acknowledgments
+We are very thankful to Jan Minx and Max Callaghan from
+the Mercator Research Institute on Global Commons and
+Climate Change (MCC) Berlin for providing us with the
+data, which is a subset of the data they used in Berrang-Ford
+et al. (2021) and Callaghan et al. (2021).
+References
+Araci, D. 2019. Finbert: Financial sentiment analy-
+sis with pre-trained language models. arXiv preprint
+arXiv:1908.10063 .
+Bender, E. M.; Gebru, T.; McMillan-Major, A.; and
+Shmitchell, S. 2021. On the Dangers of Stochastic Par-
+rots: Can Language Models Be Too Big? In Proceedings
+of the 2021 ACM Conference on Fairness, Accountability,
+and Transparency , 610–623.
+Berrang-Ford, L.; Sietsma, A. J.; Callaghan, M.; Minx, J. C.;
+Scheelbeek, P. F.; Haddaway, N. R.; Haines, A.; and Dan-
+gour, A. D. 2021. Systematic mapping of global research on
+climate and health: a machine learning review. The Lancet
+Planetary Health , 5(8): e514–e525.
+
+Bingler, J. A.; Kraus, M.; Leippold, M.; and Webersinke,
+N. 2022a. Cheap talk and cherry-picking: What C LIMATE -
+BERT has to say on corporate climate risk disclosures. Fi-
+nance Research Letters , 102776.
+Bingler, J. A.; Kraus, M.; Leippold, M.; and Webersinke, N.
+2022b. Cheap talk in corporate climate commitments: The
+role of active institutional ownership, signaling, materiality,
+and sentiment. Swiss Finance Institute Research Paper .
+Callaghan, M.; Schleussner, C.-F.; Nath, S.; Lejeune, Q.;
+Knutson, T. R.; Reichstein, M.; Hansen, G.; Theokritoff, E.;
+Andrijevic, M.; Brecha, R. J.; et al. 2021. Machine-learning-
+based evidence and attribution mapping of 100,000 climate
+impact studies. Nature climate change , 11(11): 966–972.
+Chalkidis, I.; Fergadiotis, M.; Malakasiotis, P.; Aletras, N.;
+and Androutsopoulos, I. 2020. LEGAL -BERT : The muppets
+straight out of law school. arXiv preprint arXiv:2010.02559 .
+Clark, K.; Luong, M.-T.; Le, Q. V .; and Manning, C. D.
+2020. ELECTRA: Pre-training Text Encoders as Discrimi-
+nators Rather Than Generators. In International Conference
+on Learning Representations .
+Cody, E. M.; Reagan, A. J.; Mitchell, L.; Dodds, P. S.; and
+Danforth, C. M. 2015. Climate Change Sentiment on Twit-
+ter: An Unsolicited Public Opinion Poll. PLOS ONE , 10(8):
+e0136092.
+Collobert, R.; and Weston, J. 2008. A Unified Architecture
+for Natural Language Processing: Deep Neural Networks
+with Multitask Learning. In Proceedings of the 25th Inter-
+national Conference on Machine Learning , 160–167.
+Dai, Z.; Yang, Z.; Yang, Y .; Carbonell, J. G.; Le, Q.; and
+Salakhutdinov, R. 2019. Transformer-XL: Attentive Lan-
+guage Models beyond a Fixed-Length Context. In Proceed-
+ings of the 57th Annual Meeting of the Association for Com-
+putational Linguistics , 2978–2988.
+Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2018.
+Bert: Pre-training of deep bidirectional transformers for lan-
+guage understanding. arXiv preprint arXiv:1810.04805 .
+Diggelmann, T.; Boyd-Graber, J.; Bulian, J.; Ciaramita, M.;
+and Leippold, M. 2020. CLIMATE-FEVER: A Dataset for
+Verification of Real-World Climate Claims. arXiv preprint
+arXiv:2012.00614 .
+Friederich, D.; Kaack, L. H.; Luccioni, A.; and Steffen,
+B. 2021. Automated Identification of Climate Risk Dis-
+closures in Annual Corporate Reports. arXiv preprint
+arXiv:2108.01415 .
+Gokaslan, A.; and Cohen, V . 2019. OpenWebText Corpus.
+Gr¨uning, M. 2011. Artificial intelligence measurement of
+disclosure (AIMD). European Accounting Review , 20(3):
+485–519.
+Gururangan, S.; Marasovi ´c, A.; Swayamdipta, S.; Lo, K.;
+Beltagy, I.; Downey, D.; and Smith, N. A. 2020. Don’t Stop
+Pretraining: Adapt Language Models to Domains and Tasks.
+InProceedings of the 58th Annual Meeting of the Associa-
+tion for Computational Linguistics , 8342–8360.
+Hershcovich, D.; Webersinke, N.; Kraus, M.; Bingler, J. A.;
+and Leippold, M. 2022. Towards Climate Awareness in NLP
+Research. arXiv preprint arXiv:2205.05071 .Howard, J.; and Ruder, S. 2018. Universal Language Model
+Fine-tuning for Text Classification. In Proceedings of the
+58th Annual Meeting of the Association for Computational
+Linguistics , 328–339.
+Kim, D.-Y .; and Kang, S.-W. 2018. Analysis of Recognition
+of Climate Changes using Word2Vec. International Journal
+of Pure and Applied Mathematics , 120(6): 5793–5807.
+K¨olbel, J. F.; Leippold, M.; Rillaerts, J.; and Wang, Q. 2020.
+Ask BERT: How regulatory disclosure of transition and
+physical climate risks affects the CDS term structure. Avail-
+able at SSRN 3616324 .
+Lee, J.; Yoon, W.; Kim, S.; Kim, D.; Kim, S.; So, C. H.; and
+Kang, J. 2020. BioBERT: a pre-trained biomedical language
+representation model for biomedical text mining. Bioinfor-
+matics , 36(4): 1234–1240.
+Liu, Y .; Ott, M.; Goyal, N.; Du, J.; Joshi, M.; Chen, D.; Levy,
+O.; Lewis, M.; Zettlemoyer, L.; and Stoyanov, V . 2019.
+ROBERT A: A robustly optimized bert pretraining approach.
+arXiv preprint arXiv:1907.11692 .
+Luccioni, A.; Baylor, E.; and Duchene, N. 2020. Analyzing
+Sustainability Reports Using Natural Language Processing.
+arXiv preprint arXiv:2011.08073 .
+Nagel, S. 2016. CC-NEWS.
+Rasmy, L.; Xiang, Y .; Xie, Z.; Tao, C.; and Zhi, D. 2021.
+Med-BERT: pretrained contextualized embeddings on large-
+scale structured electronic health records for disease predic-
+tion. NPJ digital medicine , 4(1): 1–13.
+Ruder, S.; and Plank, B. 2017. Learning to select data for
+transfer learning with Bayesian Optimization. In Proceed-
+ings of the 2017 Conference on Empirical Methods in Natu-
+ral Language Processing , 372–382.
+Sanh, V .; Debut, L.; Chaumond, J.; and Wolf, T. 2019. Dis-
+tilBERT, a distilled version of BERT: smaller, faster, cheaper
+and lighter. arXiv preprint arXiv:1910.01108 .
+Sautner, Z.; van Lent, L.; Vilkov, G.; and Zhang, R. 2022.
+Firm-level Climate Change Exposure. Journal of Finance ,
+forthcoming.
+Stammbach, D.; Webersinke, N.; Bingler, J. A.; Kraus, M.;
+and Leippold, M. 2022. A Dataset for Detecting Real-World
+Environmental Claims. arXiv preprint arXiv:2209.00507 .
+Trinh, T. H.; and Le, Q. V . 2019. A Simple Method for Com-
+monsense Reasoning. arXiv preprint arXiv:1806.02847 .
+Varini, F. S.; Boyd-Graber, J.; Ciaramita, M.; and Leippold,
+M. 2020. ClimaText: A dataset for climate change topic de-
+tection. In Tackling Climate Change with Machine Learning
+(Climate Change AI) workshop at NeurIPS .
+Wang, G.; Chillrud, L.; and McKeown, K. 2021. Evidence
+based Automatic Fact-Checking for Climate Change Misin-
+formation. International Workshop on Social Sensing on The
+International AAAI Conference on Web and Social Media .
+Zhu, Y .; Kiros, R.; Zemel, R.; Salakhutdinov, R.; Urtasun,
+R.; Torralba, A.; and Fidler, S. 2015. Aligning books and
+movies: Towards story-like visual explanations by watching
+movies and reading books. In Proceedings of the IEEE in-
+ternational conference on computer vision , 19–27.
+
+Appendix
+A Climate Performance Model Card
+Table 9 shows our climate performance model card, follow-
+ing Hershcovich et al. (2022).
+ClimateBert
+1. Model publicly available? Yes
+2. Time to train final model 48 hours
+3. Time for all experiments 350 hours
+4. Power of GPU and CPU 0.7 kW
+5. Location for computations Germany
+6. Energy mix at location 470 gCO2eq/kWh
+7. CO2eq for final model 15.79 kg
+8. CO2eq for all experiments 115.15 kg
+9. Average CO2eq for inference per sample 0.62 mg
+Table 9: Climate performance model card for ClimateBert.
+B Annotation Guidelines
+For our annotation procedure, we implemented the fol-
+lowing general rules. The annotators had to label climate-
+relevant paragraphs. If the paragraph was climate-relevant,
+then they had to attach a sentiment to a paragraph. Annota-
+tors were asked to apply common sense, e.g., when a given
+paragraph might not provide all the context, but the context
+might seem obvious. Moreover, annotators were informed
+that each annotation should be a 0-1 decision. Hence, if
+an annotator was 70% certain, then this was rounded up to
+100%. We asked, on average, five researchers to annotate the
+same tasks to obtain some measure of dispersion. In case of
+a close verdict or a tie between the annotators, the authors of
+this paper discussed the paragraph in depth before reaching
+an agreement.
+Text classification
+The first task was to label climate-relevant paragraphs. The
+labels are YesorNo. As a general rule, we determined that
+just discussing nature/environment can be sufficient, and
+mentioning clean energy, emissions, fossil fuels, etc., can
+also be sufficient. It is a Yes, if the paragraph includes some
+wording on a climate change or environment related topic
+(including transition and litigation risks, i.e., emission mit-
+igation measures, energy consumption and energy sources
+etc.; and physical risks, i.e., increase in risk of floods, coastal
+area exposure, storms etc.). It is a No, if the paragraph is not
+related to climate policy, climate change or an environmen-
+tal topic at all. For some examples, see Table 10.
+Sentiment Analysis
+For the sentiment analysis, annotators had to provide la-
+bels as to whether a (climate change-related) paragraph talks
+about a Risk or threat that negatively impacts an entity of in-
+terest, i.e. a company (negative sentiment), or whether an en-
+tity is referring to some Opportunity arising due to climate
+change (positive sentiment). The paragraph can also make
+just a Neutral statement.Label Examples
+Yes Sustainability: The Group is subject
+to stringent and evolving laws, reg-
+ulations, standards and best prac-
+tices in the area of sustainabil-
+ity (comprising corporate gover-
+nance, environmental management
+and climate change (specifically
+capping of emissions), health and
+safety management and social per-
+formance) which may give rise
+to increased ongoing remediation
+and/or other compliance costs and
+may adversely affect the Group’s
+business, results of operations, fi-
+nancial condition and/or prospects.
+Yes Scope 3: Optional scope that in-
+cludes indirect emissions associ-
+ated with the goods and services
+supply chain produced outside the
+organization. Included are emis-
+sions from the transport of products
+from our logistics centres to stores
+(downstream) performed by exter-
+nal logistics operators (air, land
+and sea transport) as well as the
+emissions associated with electric-
+ity consumption in franchise stores.
+No Risk and risk management Opera-
+tional risk and compliance risk Op-
+erational risk is the risk of loss re-
+sulting from inadequate or failed
+internal processes, people and sys-
+tems, or from external events in-
+cluding legal risk but excluding
+strategic and reputation risk. It also
+includes, among other things, tech-
+nology risk, model risk and out-
+sourcing risk.
+Table 10: Examples for the annotation task climate
+(Yes/No).
+To be more precise, we consider a paragraph relating to
+risk, if the paragraph mainly talks about 1) business down-
+side risks, potential losses and adverse developments detri-
+mental to the entity 2) and/or about negative impact of an
+entity’s activities on the society/environment 3) and/or asso-
+ciates specific negative adjectives to the anticipated, past or
+present developments and topics covered.
+We consider a paragraph relating to opportunities, if the
+paragraph mainly talks about 1) business opportunities aris-
+ing from mitigating climate change, from adapting to cli-
+mate change etc. which might be beneficial for a specific
+entity 2) and/or about positive impact of an entity’s activi-
+ties on the society/environment 3) and/or associates specific
+
+positive adjectives to the anticipated, past or present devel-
+opments and topics covered.
+Lastly, we consider a paragraph as neutral if it mainly
+states facts and developments 1) without putting them into
+positive or negative perspective for a specific entity and/or
+the society and/or the environment, 2) and/or does not as-
+sociate specific positive or negative adjectives to the antic-
+ipated, past or present facts stated and topics covered. For
+some examples, see Table 11.
+C Added Tokens
+’CO2’, ’emissions’, ”’, ’temperature’, ’environmental’,
+’soil’, ’increase’, ’conditions’, ’potential’, ’increased’, ’ar-
+eas’, ’degrees’, ’across’, ’systems’, ’emission’, ’precipi-
+tation’, ’impacts’, ’compared’, ’countries’, ’sustainable’,
+’provide’, ’reduction’, ’annual’, ’reduce’, ’greenhouse’,
+’approach’, ’processes’, ’factors’, ’observed’, ’renewable’,
+’temperatures’, ’distribution’, ’studies’, ’variability’, ’sig-
+nificantly’, ’–’, ’further’, ’regions’, ’addition’, ’showed’,
+’“’, ’industry’, ’consumption’, ’regional’, ’risks’, ’atmo-
+spheric’, ’supply’, ’companies’, ’plants’, ’biomass’, ’elec-
+tricity’, ’respectively’, ’activities’, ’communities’, ’cli-
+matic’, ’solar’, ’investment’, ’spatial’, ’rainfall’, ’ ’, ’sus-
+tainability’, ’costs’, ’reduced’, ’2021’, ’influence’, ’vegeta-
+tion’, ’sources’, ’possible’, ’ecosystem’, ’scenarios’, ’sum-
+mer’, ’drought’, ’structure’, ’economy’, ’considered’, ’var-
+ious’, ’atmosphere’, ’several’, ’technologies’, ’transition’,
+’assessment’, ’dioxide’, ’ocean’, ’fossil’, ’patterns’, ’waste’,
+’solutions’, ’transport’, ’strategy’, ’CH4’, ’policies’, ’un-
+derstanding’, ’concentration’, ’customers’, ’methane’, ’ap-
+plied’, ’increases’, ’estimated’, ’flood’, ’measured’, ’ther-
+mal’, ’concentrations’, ’decrease’, ’greater’, ’following’,
+’proposed’, ’trends’, ’basis’, ’provides’, ’operations’, ’dif-
+ferences’, ’hydrogen’, ’adaptation’, ’methods’, ’capture’,
+’variation’, ’reducing’, ’N2O’, ’parameters’, ’ecosystems’,
+’investigated’, ’yield’, ’strategies’, ’indicate’, ’caused’, ’dy-
+namics’, ’obtained’, ’efforts’, ’coastal’, ’become’, ’agri-
+cultural’, ’decreased’, ’GHG’, ’materials’, ’mainly’, ’rela-
+tionship’, ’ecological’, ’benefits’, ’+/-’, ’challenges’, ’nitro-
+gen’, ’forests’, ’trend’, ’estimates’, ’towards’, ’Committee’,
+’seasonal’, ’developing’, ’particular’, ’importance’, ’tropi-
+cal’, ’ratio’, ’2030’, ’composition’, ’employees’, ’charac-
+teristics’, ’scenario’, ’measurements’, ’plans’, ’fuels’, ’in-
+frastructure’, ’overall’, ’responses’, ’presented’, ’least’, ’as-
+sess’, ’diversity’, ’periods’, ’delta’, ’included’, ’already’,
+’targets’, ’achieve’, ’affect’, ’conducted’, ’operating’, ’pop-
+ulations’, ’variations’, ’studied’, ’additional’, ’construction’,
+’northern’, ’variables’, ’soils’, ’ensure’, ’recovery’, ’com-
+bined’, ’decision’, ’practices’, ’however’, ’determined’, ’re-
+sulting’, ’mitigation’, ’conservation’, ’estimate’, ’identify’,
+’observations’, ’losses’, ’productivity’, ’agreement’, ’mon-
+itoring’, ’investments’, ’pollution’, ’contribution’, ’oppor-
+tunities’, ’simulations’, ’gases’, ’statements’, ’planning’,
+’shares’, ’sediment’, ’flux’, ’requirements’, ’trees’, ’tempo-
+ral’, ’determine’, ’southern’, ’previous’, ’integrated’, ’rel-
+atively’, ’analyses’, ’means’, ’2050’, ’”’, ’uncertainty’,
+’pandemic’, ’fluxes’, ’findings’, ’moisture’, ’consistent’,
+’decades’, ’snow���, ’performed’, ’contribute’, ’crisis’Label Examples
+Opportunity Grid & Infrastructure and Retail – today represent
+the energy world of tomorrow. We rank among Eu-
+rope‘s market leaders in the grid and retail busi-
+ness and have leading positions in renewables. We
+intend to spend a total of between Euro 6.5 bil-
+lion and Euro 7.0 billion in capital throughout the
+Group from 2017 to 2019.
+Opportunity We want to contribute to the transition to a circu-
+lar economy. The linear economy is not sustain-
+able. We discard a great deal (waste and there-
+fore raw materials, experience, social capital and
+knowledge) and are squandering value as a result.
+This is not tenable from an economic and ecolog-
+ical perspective. As investor we can ‘direct’ com-
+panies and with our network, our scale and our in-
+fluence we can help the movement towards a cir-
+cular future (creating a sustainable society) further
+along.
+Neutral A similar approach could be used for allocating
+emissions in the fossil fuel electricity supply chain
+between coal miners, transporters and generators.
+We don’t invest in fossil fuel companies, but those
+investors who do should account properly for their
+role in the production of dangerous emissions from
+burning fossil fuels.
+Neutral Omissions: Emissions associated with joint ven-
+tures and investments are not included in the emis-
+sions disclosure as they fall outside the scope of our
+operational boundary. We do not have any emis-
+sions associated with heat, steam or cooling. We
+are not aware of any other material sources of omis-
+sions from our emissions reporting.
+Risk We estimated that between 36.5 and 52.9 per cent
+of loans granted to our clients are exposed to tran-
+sition risks. If the regulator decides to pass am-
+bitious laws to accelerate the transition towards a
+low-carbon economy, carbon-intensive companies
+would incur in higher costs, which may prevent
+them from repaying their debt. In turn, this would
+weaken our bank’s balance sheets. .
+Risk American National Insurance Company recognizes
+that increased claims activity resulting from catas-
+trophic events, whether natural or man-made, may
+result in significant losses, and that climate change
+may also affect the affordability and availability of
+property and casualty insurance and the pricing for
+such products.
+Table 11: Examples for the annotation task sentiment (Op-
+portunity/Neutral/Risk).
+

aaaifss2022_13.txt

@@ -0,0 +1,488 @@
+Curator: Creating Large-Scale Curated Labelled Datasets using Self-Supervised
+Learning
+Tarun Narayanan*1, Ajay Krishnan*1, Anirudh Koul1,2,4, Siddha Ganju1,3,4
+1SpaceML2Pinterest3NVIDIA4Frontier Development Lab
+[email protected], [email protected]
+Abstract
+Applying Machine learning to domains like Earth Sciences is
+impeded by the lack of labeled data, despite a large corpus
+of raw data available in such domains. For instance, train-
+ing a wildfire classifier on satellite imagery requires curat-
+ing a massive and diverse dataset, which is an expensive and
+time-consuming process that can span from weeks to months.
+Searching for relevant examples in over 40 petabytes of un-
+labelled data requires researchers to manually hunt for such
+images, much like finding a needle in a haystack. We present
+a no-code end-to-end pipeline, Curator1, which dramatically
+minimizes the time taken to curate an exhaustive labeled
+dataset. Curator is able to search massive amounts of unla-
+belled data by combining self-supervision, scalable nearest
+neighbor search, and active learning to learn and differentiate
+image representations. The pipeline can also be readily ap-
+plied to solve problems across different domains. Overall, the
+pipeline makes it practical for researchers to go from just one
+reference image to a comprehensive dataset in a diminutive
+span of time.
+Introduction
+One of the initial steps for a scientific study related to cli-
+mate change and natural disasters, including wildfires, oil
+spills, hurricanes, dust storms, etc., involves scientists gath-
+ering a large number of relevant examples from satellite
+imagery. Locating an exhaustive set of examples requires
+painstakingly inspecting 197 million square miles of satel-
+lite imagery each day across more than 20 years. While such
+an effort can produce a valuable trove of data, the act of man-
+ually searching is laborious, expensive, and often imprac-
+tical - grounding many scientific studies before they could
+ever take off.
+While one of the approaches to solving this is building an
+image similarity search, several challenges arise when ap-
+plying similarity search to raw satellite imagery:
+• The data is unlabelled, preventing attempts to train con-
+ventional supervised models which could have generated
+meaningful representations.
+*These authors contributed equally.
+Copyright © 2022, Association for the Advancement of Artificial
+Intelligence (www.aaai.org). All rights reserved.
+1We release all instructions, trained models and code for Cura-
+tor : https://www.github.com/spaceml-org• Pretrained ImageNet (Deng et al. 2009) models fail to
+transfer representations and generalize to this data - es-
+pecially for larger areas that are usually without sharp
+edges including clouds as well as multi-spectral data.
+• Climate phenomena can have vastly different physical
+sizes - from few miles for wildfires to 300+ miles for
+hurricanes.
+• Vast data imbalances inherently present in the data.
+• The engineering challenges that come with the sheer
+scale of our data.
+We propose Curator, a modular toolkit that aims to take
+a user from one reference image to an exhaustive set of rel-
+evant examples for any large unlabelled image data source.
+It solves the core issue of data inaccessibility by discovering
+relevant samples from sizeable collections while minimizing
+human labeling effort. This pipeline combines several indi-
+vidually tested, high-performance components built for spe-
+cific tasks - from downloading data, training self-supervised
+models, large-scale similarity search, active learning, and
+crowd-sourced labeling. This open source project, built by
+citizen scientists (Koul et al. 2020), aims to enable a re-
+searcher to accomplish all of this without writing a single
+line of code or possessing any prerequisite AI knowledge.
+This ease of usage further reduces barriers to entry and hope-
+fully catalyzes research involving climate science.
+Previous Methods
+We demonstrate a specific use case of our pipeline that aims
+to solve a previously unsolved problem - building a curated
+dataset for any natural phenomenon by intelligently index-
+ing over 897 satellite imagery sources via the Global Im-
+agery Browse Services (GIBS) portal. To the best of our
+knowledge, the widely accepted solution in practice which
+also acts as the baseline is a manual approach involving vi-
+sual inspection of data from the GIBS portal for multiple
+layers over a region for a period of time, where each layer
+provides information overlays based on science disciplines,
+hazards, and disaster categories, downloading the requisite
+data, and then manually annotating it. Research in this field
+includes manually labeling, semi-supervised learning like in
+(Kim et al. 2019), or using text mining and NLP techniques
+to extract images and their labels from multiple large data
+stores. Our method involves annotating a negligible number
+
+Figure 1: The Curator pipeline
+of images in comparison, and then relies on active learning
+to generate weak labels for the rest of the dataset.
+Pipeline
+Our key goal is to let a scientist use a single query image (say
+of a climate event) to ultimately identify every potential ex-
+ample of the same category in a large image collection (like
+satellite imagery). A scalable way to do this is by evaluating
+each image with a classifier tuned to the user’s needs. Train-
+ing such a supervised classifier requires enough positive and
+negative examples for training. Getting to this training set
+can be achieved in four steps - (1)training a self-supervised
+model on unlabeled data, in order to learn semantically rel-
+evant representations. (2)generating embeddings for the en-
+tire dataset (3)for one or more starter examples, building a
+seed set of similar images, i.e images with embeddings cor-
+responding to the nearest neighbours to the query image (4)
+using several iterations of human-in-the-loop active learning
+to find examples that maximize classifier performance while
+minimizing human labeling time.
+The modules of Curator can be combined to achieve this
+functionality (summarized in Fig 1). Key themes in their de-
+velopment include that each tool need to be 1) executable
+through a single command 2) highly modular so it can be
+used for an individual task or combined for a range of tasks,
+including beyond climate science 3) built for high perfor-
+mance with the available hardware (single, multiple GPU or
+multi node) while being cost effective at scale. With the aim
+to get researchers started in minutes, the tools can be run
+on a local machine through a simple command line inter-
+face. For higher scale, the pipeline provides a cloud specific
+template using Google Cloud (which can be replicated but
+needs deeper familiarity with the cloud). We also include
+a set of data preprocessing functions that were designed tosolve some inherent deficiencies present in the satellite im-
+agery data (for more information see Appendix A).
+GIBS Downloader
+GIBS Downloader (Lisboa et al. 2021) is a command-line
+tool that simplifies access to satellite imagery from NASA
+Global Imagery Browse Services (GIBS), thereby tackling
+all the esoteric challenges behind acquiring and processing
+decades of satellite imagery data. It provides access to over
+897 products, along with the ability to search their remote
+sensing product descriptions by keywords. It offers vari-
+ous functionalities to easily convert datasets to a format that
+can be directly used for AI training, including TensorFlow’s
+TFRecords for accelerating the speed of data ingestion in
+training pipelines. The required arguments include the date
+range and the lat/long coordinates of the rectangular region.
+Operating on a canvas of up to 262144 x 131072 pixels for
+a full view of the globe (which cannot be opened by most
+image viewers), it uses several performance optimizations
+like multithreading to parallelize extraction of smaller tiles
+suited for a researcher’s needs.
+Self Supervised Learner
+Self Supervised Learner is a command-line tool that takes
+a directory of unlabeled images and trains a self-supervised
+model. Self-Supervised Learning (SSL) is a relatively new
+method of unsupervised representation learning wherein we
+generate temporary labels intrinsically from the images by
+exposing a relationship between different parts of the image
+or with multiple views of the image. Currently, the SimCLR
+(Chen et al. 2020), and the SimSiam (Chen and He 2021)
+architectures are supported. Built for performance, the Self-
+Supervised Learner utilizes NVIDIA DALI package to par-
+allelize CPU operations like image decoding and augmenta-
+
+Active Learning Strategy F1
+Score
+(Val)Total
+labelling
+effort by
+the userPositive
+Images
+RetrievedFalse
+Positive
+Images
+Retrieved
+Random Sampling (with Imagenet Pretraining) 0.45 7.6% 65% 37%
+Uncertainty Sampling (with SSL Pretraining) 0.74 7.8% 88% 12%
+Table 1: Number of positive images along with the percentage of data predicted as False Positives, that were retrieved across
+different active learning strategies.
+tions on the GPU, resulting in up to 8x speedup in training
+time. The tool can scale training from single GPU to multi
+GPU, consistently with 90% GPU resource utilization off-
+the-shelf. It also provides a high level of customizability in
+defining custom model architectures, augmentations, along
+with planned support for multi-band data and seasonal con-
+trast modeling (Ma ˜nas et al. 2021).
+Scalable Image Search
+Curator provides a command-line tool for local machines as
+well as a Google Cloud template to perform scalable interac-
+tive image search. First, the Image Embedding Indexer takes
+a model and generates embeddings rapidly (through GPU
+acceleration using NVIDIA DALI). Then, these embeddings
+are indexed for fast approximate nearest neighbor search us-
+ing FAISS (Cheng, Han, and Lu 2017). Lastly, a low latency
+API provides image query capabilities along with filtering
+options. Additionally, the pipeline provides an interactive UI
+to visualize search results. The search index is partitioned
+by date, resolution, and product to make the system scal-
+able and parallelizable. For an image collection with up to
+5 million images, most modern laptops can retrieve results
+in under a second, satisfying the requirements of most re-
+searchers and enabling them to get started quickly. For larger
+collections, the cloud template contains several performance
+tweaks to parallelize and run a scalable yet cost-efficient
+multi-node system, such as utilizing Google Google Cloud
+Functions, reading the index files as a byte stream, configur-
+ing the same regions for bucket and VM regions, and more.
+Swipe Labeler
+Swipe Labeler is a browser-based annotation tool meant to
+quickly and efficiently label image collections with binary
+labels. It is intended to make the usually tedious process
+of labeling data more engaging by swiping right or left (or
+pressing right/left arrow keys) to move the images into fold-
+ers categorized as relevant and non-relevant. Accessible on
+both mobile and desktop, the tool can be activated by a sin-
+gle command. The tool offers multi-user collaborative label-
+ing by seamlessly generating a public shareable link without
+the user requiring any networking knowledge.
+Active Labeler
+Active Labeler (AL) is a tool that incorporates human-in-
+the-loop active learning to minimize labeling while maxi-
+mizing classifier performance. Given a seed set of labeled
+images, it trains a classifier (transfer learning on the SSLmodel, or any image classification setup), evaluates all un-
+labeled images and picks a small subset for human labeling,
+which are added to the labeled image set. It repeats this pro-
+cess iteratively till the classifier shows robust performance
+metrics. A variety of strategies can be employed to identify
+the data points that would contribute most to the accuracy of
+the model, in other words, they calculate which data points
+are most ’influential’. The tool supports a range of strategies
+fundamentally based on Uncertainty Sampling such as Least
+Confidence Sampling, Margin Based Sampling, and Entropy
+Based Sampling. With a sampling strategy that is based only
+on uncertainty, there is a possibility that the samples selected
+for training are very similar to each other. In such a scenario,
+intuitively, the model would only learn about a certain type
+of image in each iteration, rendering the process inefficient.
+The inclusion of diversifying sampling strategies may help
+fully utilize each iteration, ensuring that the model learns
+from a set of diverse samples as opposed to a homogeneous
+one. The strategies that have been implemented thus far are
+Iterative proximity-based sampling, Gaussian Sampling and
+Clustering-based sampling. Beyond basic active learning,
+AL also interfaces with Scalable Image Search. It helps build
+a labeled seed set by taking a single starter image, retriev-
+ing similar images, and labeling them with Swipe Labeler.
+The seed images should contain distinguishable features that
+you want to distinctively see in the retrieved similar images.
+At scale, several performance tweaks have been incorpo-
+rated, including - (1)using embeddings instead of images
+to significantly reduce computation (2)training a classifica-
+tion head using features from a pretrained SSL backbone (3)
+reducing the output dimension of the SSL backbone, to im-
+prove downstream training time and space efficiency. Addi-
+tionally, we utilize a subsample of approximately equidistant
+embedding vectors (Core-Set) instead of the entire embed-
+ding space in order to exponentially reduce the time taken to
+perform a forward pass operation (for more, refer Appendix
+A). The datapoints selected in the subsample are then used
+to find the nearest neighbors in the entire embedding space.
+With these improvements, leveraging multi-million to bil-
+lion scale image datasets becomes practical from a cost and
+latency standpoint.
+Results
+To evaluate the effectiveness of the pipeline on a labeled
+benchmark dataset containing satellite imagery, we exper-
+imented with RESISC45 (Cheng, Han, and Lu 2017) (Re-
+mote Sensing Image Scene Classification), which contains
+
+Figure 2: Image Retrieval results on VIIRS data. (Left) Query Image (Right) Retrieved images from the curated set.
+31,500 images, covering 45 classes with 700 images in each
+class with high intra-class diversity and inter-class similar-
+ity, making it relatively challenging. Given a single refer-
+ence image, we aim to evaluate the number of images of
+the same class that can be identified, along with the amount
+of human labeling required. For a starter image, a seed set
+is constructed and then assigned positive/negative class la-
+bels. This seed set consists of 64 nearest neighbors to the
+starter image and 32 randomly sampled images to provide a
+diverse negative class. This seed set is used by the Active La-
+beler, which iteratively trains a classifier, classifies the entire
+dataset, and picks a subset of 64 most informative images to
+be assigned a label, which is then used in the subsequent iter-
+ation for training. The system runs till 5% of the dataset has
+been labeled. The resulting classifier is then used to identify
+potential positive classes in the dataset and presented to the
+user for verification to build a curated set. We repeat the ex-
+periment for all 45 classes, with 10 randomly chosen starter
+images per class. Results, shown in Table 1, showcase that,
+on average, 88% of the images belonging to the same class
+as the starter image was retrieved with 7.8% manual labeling
+effort. This result is in contrast to the baseline of manually
+evaluating every single image in the dataset.
+To further battle test our pipeline in real-time data sce-
+narios, we setup Curator to curate images from an unla-
+beled satellite imagery dataset. We tiled and retrieved 30
+days’ worth of data from the VIIRS product using the GIBS
+Downloader tool, and we pretrained SimCLR on this data
+using relevant augmentation strategies for 1000 epochs on
+a single GPU. This model is the backbone for Active La-
+beler. We picked starter images from our validation set andpassed them to Curator to retrieve similar images. Examples
+of starter images and images from their curated set are illus-
+trated in Figure 2.
+We believe another important outcome of using our
+pipeline is the underlying time and monetary benefit that
+comes from rapid iteration. For example, let’s evaluate the
+task of finding images of islands from NASA Worldview.
+During a recent demonstration of Curator on the NASA
+GIBS/Worldview imagery pipeline, a machine was trained
+to search for islands through five million tiles of Earth im-
+agery starting with a single seed image of an island. Ap-
+proximately 1,000 islands were identified in just 52 minutes
+with just one human in the loop. If done manually, this effort
+would take an estimated 7,000 hours (assuming five seconds
+to evaluate and label each image tile) and potentially cost
+as much as $105,000 (assuming $15 per hour per annotator)
+(Blumenfeld 2021).
+Conclusion
+We present a novel pipeline that provides an automated ap-
+proach to curating relevant datasets starting from a single
+image with significantly less human effort involved. Built
+for scale and cost effectiveness, the pipeline leverages tech-
+niques like self-supervised learning, human-in-the-loop ac-
+tive learning, geometric data sampling, and nearest neighbor
+search. Reducing the time of manual data curation from sev-
+eral months to hours or even minutes opens new avenues of
+scientific exploration previously considered impractical. By
+releasing a readily usable open-source toolbox, we hope to
+accelerate research in domains like climate science, where
+access to structured data and has been a major challenge.
+
+References
+Blumenfeld, J. 2021. SpaceML: Rise of the Machine (Learn-
+ing).
+Chen, S.; Cao, E.; Koul, A.; Ganju, S.; Praveen, S.; and
+Kasam, M. A. 2021. Reducing Effects of Swath Gaps on
+Unsupervised Machine Learning Models for NASA MODIS
+Instruments. arXiv preprint arXiv:2106.07113 .
+Chen, T.; Kornblith, S.; Norouzi, M.; and Hinton, G. 2020.
+A simple framework for contrastive learning of visual repre-
+sentations. In International conference on machine learning ,
+1597–1607. PMLR.
+Chen, X.; and He, K. 2021. Exploring simple siamese repre-
+sentation learning. In Proceedings of the IEEE/CVF Confer-
+ence on Computer Vision and Pattern Recognition , 15750–
+15758.
+Cheng, G.; Han, J.; and Lu, X. 2017. Remote Sensing Im-
+age Scene Classification: Benchmark and State of the Art.
+CoRR , abs/1703.00121.
+Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; and Fei-
+Fei, L. 2009. ImageNet: A large-scale hierarchical image
+database. In 2009 IEEE Conference on Computer Vision
+and Pattern Recognition , 248–255.
+Kim, T. K.; Yi, P. H.; Hager, G. D.; and Lin, C. T. 2019.
+Refining dataset curation methods for deep learning-based
+automated tuberculosis screening. Journal of Thoracic Dis-
+ease, 12(9).
+Koul, A.; Ganju, S.; Kasam, M.; and Parr, J. 2020. Space
+ML: Distributed Open-source Research with Citizen Scien-
+tists for the Advancement of Space Technology for NASA.
+CoRR , abs/2012.10610.
+Lisboa, F.; Verma, S.; Koul, A.; Kasam, M. A.; and Ganju, S.
+2021. Democratizing Earth Science Research with Accessi-
+ble Data High-Performance Training Pipelines. Committee
+on Space Research Cloud Computing Workshop .
+Ma˜nas, O.; Lacoste, A.; Giro-i Nieto, X.; Vazquez, D.;
+and Rodriguez, P. 2021. Seasonal Contrast: Unsupervised
+Pre-Training from Uncurated Remote Sensing Data. arXiv
+preprint arXiv:2103.16607 .
+Appendix A: Adapting to Different Tasks
+The pipeline is generalizable on any unlabelled source of
+data in a domain agnostic manner. Performing this task sim-
+ply requires us to define a custom Data Source, Data Down-
+loader, and an optional Data Preprocessor that is specific to
+the problem we’re solving.
+Data Source
+The Data Source is a user-provided pool of unlabelled data.
+Most domains have a lot of data being collected that cur-
+rently do not translate to value in our context due to their
+lack of organization, and Curator is designed to leverage
+these data sources without the hard requirement for anno-
+tation. In our demonstration we pick the NASA Worldview
+platform as our Data Source, and we demonstrate how our
+pipeline can be used to generate curated datasets from the
+satellite imagery data available on this platform.
+Figure 3: Trained convolutional autoencoder outputs for
+Swath Filler. Query image (leftmost column) and its corre-
+sponding most-similar four images. Filling strategy changes
+row wise: no fill, Random RGB, Pixel RGB, Neighbor RGB.
+Random RGB fill strategy results show that the autoencoder
+focuses on swath gap positions. Neighbor RGB fill strategy
+results show that the autoencoder ignores the swath gap and
+concentrates on the ROI.
+Data Downloader
+Our data source can be a vast stream of unlabelled images,
+but that data cannot be directly used to train machine learn-
+ing models due to the lack of compute and storage. Frame-
+works also require datasets to adhere to a specified format.
+The Data Downloader helps source limited data from the
+data source and converts it into a format that can be directly
+used by the model training framework. Curator allows full
+flexibility to the user in defining the Data Source and the
+Data Downloader based on their domain.
+Data Preprocessor
+Data preprocessing is a fundamental data operation in ML
+that helps improve the model’s performance. Data present
+in certain domains like satellite imagery, medical imaging,
+and the like come with inherent discrepancies. Data Pre-
+processor consists of a set of statistical and geometric func-
+tions that were designed to solve some inherent deficiencies
+present in the satellite imagery data. These challenges are
+specific to the dataset. For instance, the NASA Worldview
+data had some esoteric deficiencies that we had to fix in or-
+der to make the data usable.
+Cloud Removal Clouds are a major barrier in Remote
+Sensing datasets since they occlude the information of the
+space underneath. They make learning representations much
+harder for Machine Learning models. We were able to re-
+trieve a cloudless version of an area by performing Image
+Subtraction over multiple images of the same area across
+several days. Contrarily, we were also able to retrieve cloud
+masks out of images individually, which can greatly help
+with cloud segmentation problems (see Figure 4(a)).
+
+(a) Left: Gulf of Mexico without Clouds generated based on previously available data.
+Right: Generated cloud masks over the Alps region.
+(b) Image retrieval across multiple resolutions for our Tile-based multi-resolution search
+against Image-based multi-resolution search
+Figure 4: Cloud Removal and Multi-Resolution Image Search
+Swath-Fillers Image tiles retrieved from the Worldview
+MODIS product come with small gaps at the equator, called
+Swaths. These occur due to the nature of the movement of
+the satellite over the earth. Training models on images con-
+taining Swaths meant an ML model learns this as a feature
+across images and clusters them together. This affects per-
+formance greatly. Through the Nearest pixel interpolation
+strategy, we were able to perform a Content-Aware fill on
+these swaths with relevant surrounding information (Chen
+et al. 2021) (see Figure 3). This problem has also recently
+been overcome by sourcing our data from another product
+named VIIRS on GIBS.
+Multi-Resolution Image Search Images in Remote Sens-
+ing datasets can appear in different resolutions. There can be
+images with the class object appearing in different sizes, as
+well as the presence of multiple objects in an image. Similar-
+ity search precision can be affected due to this. By tiling the
+image into a grid of patches, and obtaining the nearest neigh-
+bors for each of those tiles, we were able to aggregate the
+results by using a bucket voting strategy. This helped put the
+embedding distances into context and return similar matches
+to the entire image based on the voted scoring(See Figure
+4(b)). Although in practice, we found that Multi-Resolution
+search was a time consuming process that struggled at scale,
+so instead we built a model store that consists of models
+trained on multiple resolutions. We utilize the correspond-
+ing model based on the resolution of the image being used.Diverse Data Sampler Data Imbalance is a real problem
+in Machine Learning. For instance, satellite imagery datasets
+are inherently biased due to the natural imbalance between
+the different classes present in them. 71% of the tiles present
+consist of water bodies, and our ML systems find it hard to
+learn information about poorly represented classes such as
+those images of natural phenomena, due to their sheer lack
+of occurrence in the data.
+We apply a coreset strategy to the data to obtain a more
+representative sample of our data. This was absolutely nec-
+essary since we had the resources to only train on a subset of
+our entire pool of satellite imagery data. In simpler terms, we
+pick the farthest point for the current set of points, until the
+set equals the sample size. The resulting embedding space
+is an equidistant set of points that represent a diverse sub-
+set. This diverse subset is believed to contribute more infor-
+mation to a model during training compared to a randomly
+sampled subset. The standard algorithm is a deterministic
+operation for a given starting point, and the number of oper-
+ations done is subset size * subset size * total num samples
+For a more scalable version, we also introduce a stratified
+version of this sampler,where instead of going through the
+entire embedding space, this technique first samples a ran-
+dom set of points, determines the farthest point from that
+sample, resamples a new set of points and repeats the pro-
+cess until a diverse sample is obtained. Resampling is done
+periodically to prevent the selection of farthest points within
+a sample of the embedding space. Num operations done is
+subset size * subset size * num random samples
+
+While working with large scale satellite datasets, like the
+one from NASA Worldview, we found that it was extremely
+time consuming to perform a forward pass over all 10 mil-
+lion tiled images from the dataset. Instead we employed the
+Diverse Data Sampler to pick a highly representative sample
+of just 10% of the data, thereby significantly reducing the
+time taken to perform a forward pass. Overall, along with
+the aforementioned optimizations, there is potential to re-
+duce the runtime from initally taking 21,000 hours to just 13
+minutes with no degradation in model quality.
+

aaaifss2022_14.txt

@@ -0,0 +1,495 @@
+De-risking Carbon Capture and Sequestration with Explainable CO 2Leakage
+Detection in Time-lapse Seismic Monitoring Images
+Huseyin Tuna Erdinc,*1Abhinav Prakash Gahlot,*2Ziyi Yin,3
+Mathias Louboutin,2Felix J. Herrmann,1,2,3
+1School of Electrical and Computer Engineering, Georgia Institute of Technology
+2School of Earth and Atmospheric Sciences, Georgia Institute of Technology
+3School of Computational Science and Engineering, Georgia Institute of Technology
+{herdinc3, agahlot8, ziyi.yin, mlouboutin3, felix.herrmann }@gatech.edu,
+Abstract
+With the growing global deployment of carbon capture and
+sequestration technology to combat climate change, monitor-
+ing and detection of potential CO 2leakage through existing
+or storage induced faults are critical to the safe and long-term
+viability of the technology. Recent work on time-lapse seis-
+mic monitoring of CO 2storage has shown promising results
+in its ability to monitor the growth of the CO 2plume from
+surface recorded seismic data. However, due to the low sen-
+sitivity of seismic imaging to CO 2concentration, additional
+developments are required to efficiently interpret the seis-
+mic images for leakage. In this work, we introduce a binary
+classification of time-lapse seismic images to delineate CO 2
+plumes (leakage) using state-of-the-art deep learning models.
+Additionally, we localize the leakage region of CO 2plumes
+by leveraging Class Activation Mapping methods.
+Introduction
+According to the International Energy Agency and the In-
+ternational Panel on Climate Change report (IPCC 2018),
+there is a need for a 50 percent reduction of greenhouse
+gas emissions by 2050 to avoid an increase of 1.5 degrees
+Celsius of Earth’s average temperature. This can only be
+achieved by reduced dependence on fossil fuels, use of re-
+newable sources of energy and large-scale global deploy-
+ment of carbon reduction technologies such as carbon cap-
+ture and sequestration (CCS). This technology consists of
+collection, transportation, and injection of CO 2into an ap-
+propriate geologic storage reservoir for extended time peri-
+ods (tens of years). Especially, unlike other solutions, CCS
+is considered a relatively low-cost, long-term and imminent
+solution. However, potential CO 2leakage from the under-
+ground reservoirs due to pre-existing or pressure-induced
+faults poses risks (Ringrose 2020). Thus, it is necessary to
+de-risk CCS projects by monitoring CO 2plumes in order to
+accurately detect and predict potential leakages as early as
+possible.
+Time-lapse seismic monitoring has been introduced as
+a reliable technology to monitor the CO 2dynamics in
+the Earth’s subsurface during carbon sequestration (Lumley
+*These authors contributed equally.
+Copyright © 2022, Association for the Advancement of Artificial
+Intelligence (www.aaai.org). All rights reserved.2001) and is already in use at existing storage sites (Arts
+et al. 2008; Chadwick et al. 2010; Ringrose et al. 2013; Furre
+et al. 2017). In essence, sequential (i.e once every 6 month-
+s/year/...) seismic datasets, called vintages, are collected in
+the field over an area covering the storage reservoir. Then,
+each seismic dataset is inverted to obtain high fidelity im-
+ages of the subsurface over time (Arts et al. 2008; Ayeni
+and Biondi 2010; Yin, Louboutin, and Herrmann 2021). The
+evolution of the CO 2reservoir can finally be visualized by
+subtracting the seismic images between different points in
+time. However, due to the inherently weak and noisy ampli-
+tudes of the CO 2reservoir’s response in those seismic im-
+ages, detecting the presence of potential irregularities, such
+as in CO 2plumes, corresponding to a leakage is a challeng-
+ing problem. To tackle this difficulty, we propose a machine
+learning based detection method based on standard binary
+classification.
+Recently, numerous methods leveraging machine learn-
+ing have been introduced for the detection of CO 2leakage
+based on a simple artificial neural network (ANN) (Li et al.
+2018a), and a combination of convolutional neural networks
+(CNN) and Long Short-Term Memory (LSTM) networks
+(Zhou et al. 2019). While leading to accurate predictions,
+these methods usually rely solely on the field recorded data
+rather than the subsurface seismic images. Besides, practical
+considerations such as repeatability (the ability to record the
+data in the exact same way every year) hinders their appli-
+cability to real world cases. On the other hand, as we rely
+on visualizing the CO 2plumes in the seismic image, we
+can take advantage of advanced seismic imaging techniques
+designed for non-repeated seismic acquisition such as the
+joint recovery model (JRM) (Oghenekohwo and Herrmann
+2017a; Wason, Oghenekohwo, and Herrmann 2017; Yin,
+Louboutin, and Herrmann 2021). Additionally, this imag-
+ing technique has demonstrated higher fidelity imaging than
+sequential seismic imaging allowing for easier detection of
+CO2leakage.
+We will show in the following sections that we can effi-
+ciently and accurately detect CO 2from realistic seismic im-
+ages recovered by JRM on synthetic but representative mod-
+els of the Earth subsurface. We demonstrate our method us-
+ing different state-of-the-art deep learning models in a trans-
+fer learning setting to classify CO 2plume seismic images
+
+with regular (no-leakage) CO 2plume or with CO 2leakage.
+As CO 2leakage detection needs trustworthiness, we further
+unravel the decisions made by our models and utilize Class
+Activation Mapping (CAM) methods (Zhou et al. 2015) to
+identify and visualize seismic image areas crucial for model
+classification results. We show that the CAM result accu-
+rately focuses on the leakage portion of the CO 2plume and
+reservoir, validating that our network detects leakage based
+on state of the CO 2reservoir over time.
+Our main contributions are organized as follows. First, we
+introduce the classification models used for leakage detec-
+tion and the CAM methods for visualizing the area of inter-
+est in the seismic image. Second, we demonstrate the accu-
+racy of our models and qualitatively examine the results of
+our CAM methods on a realistic synthetic set of CO 2plume
+images.
+Methodology
+In order to speed up the training process and to compensate
+for the overfitting that may occur with modest sized datasets,
+we rely on transfer learning (Yosinski et al. 2014) using pre-
+trained state-of-the-art models as a starting point. In particu-
+lar, we consider four modern architectures known to achieve
+high accuracy on standard dataset such as ImageNet-1k
+(Russakovsky et al. 2015). The models used are VGG (Si-
+monyan and Zisserman 2014), ResNet (He et al. 2016), Vi-
+sion Transformer (ViT) (Dosovitskiy et al. 2021), and Swin
+Transformer (Swin) (Liu et al. 2021), all pre-trained on the
+standardized ImageNet-1k dataset.
+VGG: is a convolutional neural network (CNN) model
+that achieved significant success in The ImageNet Large
+Scale Visual Recognition Challenge (ILSVRC) competition
+in 2014 (Simonyan and Zisserman 2014). VGG consists of
+sequences of convolution and maxpool layers. In our numer-
+ical experiments, the VGG16 variant with 16 trainable layers
+is used.
+ResNet: is a CNN architecture with residual connections
+proposed to solve the vanishing gradient problem in very
+deep networks (He et al. 2016). ResNet consists of resid-
+ual blocks and each residual block has convolution layers
+and shortcut connections performing identity mapping. In
+our numerical experiments, the ResNet34 variant with 34
+trainable layers is used.
+ViT: is an architecture based on transformer which is used
+in the field of Natural Language Processing (NLP) (Vaswani
+et al. 2017). Internally, the transformer learns a relationship
+between input token pairs, and uses 16x16 patches of im-
+ages as input tokens (Dosovitskiy et al. 2021). In our numer-
+ical experiments, the tiny ViT variant is used allowing lower
+memory and computational imprint.
+Swin: is a special type of ViT that represents image
+patches hierarchically by starting from small-sized patches
+and gradually increasing the size through merging to achieve
+scale-invariance property (Liu et al. 2021). Compared to
+ViT, Swin transformer has superior (linear) computational
+efficiency by computing self-attention within certain patches
+of a window. In our numerical experiments, tiny Swin vari-
+ant is used allowing lower memory and computational im-
+print.Hyperparameters VGG16 ResNet34 ViT Swin
+Batch Size 8 8 8 8
+Learning Rate 5x10−56x10−34x10−310−3
+Exp Decay Rate( γ)0.95 0 .92 0 .98 0 .98
+Table 1: Training hyperparameters for the four models. All models
+were trained with the same number of epochs and optimizer.
+CAM Methods
+Deep learning models for classification are notoriously
+treated as “black boxes” as they do not expose their inter-
+nal knowledge or operations to its users and do not pro-
+vide interpretable results. To solve this problem, CAM based
+saliency maps (heatmaps) were introduced to highlight the
+most class-discriminative regions of to-be-classified input
+images (Zhou et al. 2015). Since CO 2leakage requires high
+fidelity, transparent and interpretable models, we use CAM
+to further make our model results explainable and highlight
+the regions of the seismic image that are most relevant to
+the classification results. In our study, we considered two
+CAM methods. First, Grad-CAM (Selvaraju et al. 2019), a
+gradient-based CAM method considered as the state-of-the-
+art in terms of explainability of neural networks for classi-
+fication. This CAM method extracts gradients from a spe-
+cific layer of a model and computes the weighted aver-
+age of that specific layer’s activations. Second, we consider
+Score-CAM (Wang et al. 2020), a perturbation based CAM
+method. Score-CAM also computes the weighted average of
+activations of a user-specified layer but, unlike Grad-CAM,
+Score-CAM relies on propagating (forward pass through the
+network) a masked input image where the mask is obtained
+via upsampling the activations of the user-defined layer.
+This CAM method provides high accuracy and interpretable
+heatmaps and alleviates potential noise and spread present
+in the gradient used for the Grad-CAM heatmaps.
+Numerical Case Study
+To generate the training dataset of CO 2plume evolution, we
+used five 2D vertical slices extracted from the 3D Compass
+velocity model (E. Jones et al. 2012) shown in Fig. 1(a). This
+model is a synthetic but realistic model representative of the
+complex geology of the southeast of the North Sea. The di-
+mension of each model (slice) used in our work is 2131 X
+4062 m2. We used FwiFlow (Li et al. 2020), to simulate the
+CO2flow dynamics and JUDI (Witte et al. 2019) to model
+the seismic data and compute the seismic images of the sub-
+surface.
+Time-lapse reservoir and seismic simulation
+We consider a realistic two well setting with a fixed injection
+well injecting CO 2and a production well extracting brine
+from subsurface storage reservoir. Injection of supercriti-
+cal CO 2into saline aquifers is an example of multi-phase
+flow in porous media. While we understand more compli-
+cated geothermal, geochemical and geomechanical process
+may eventually be considered to model the CO 2dynamics,
+we follow the two-phase immiscible incompressible flow
+
+Figure 1: Workflow for CO 2Leakage Monitoring
+physics, which in its leading order describes the process
+of supercritical CO 2displacing brine in the pore space of
+the rock. The system is governed by conservation of mass
+and Darcy’s law. We refer to the existing literature (Li et al.
+2020; Wen, Tang, and Benson 2021) (Li et al. 2020) for more
+details about this physical system.
+Using empirical relation and the Kozeny-Carman equa-
+tion(Costa 2006), the acoustic properties (velocity and den-
+sity) from the Compass model were converted into perme-
+ability and porosity (Fig. 1(b)) to simulate the multi-phase
+flow (CO 2and brine in porous media) in the reservoir. We
+used FwiFlow.jl (Li et al. 2020) to solve multi-phase flow
+equations based on the finite volume method. We simulated
+the CO 2flow for a duration varying between 7to12years
+(Fig. 1(c)). The reservoir was initially filled with saline wa-
+ter and we injected compressed CO 2at the rate of 1MT/-
+day into the reservoir for all simulations. In order to mimic
+CO2leakage, we then created a fracture at a random location
+along the top seal of the reservoir when the pressure induced
+by the CO 2injection reaches a threshold of 15MPa. We
+then converted back these simulated CO 2saturation snap-
+shots over time into wave properties with the patchy sat-
+uration model (Avseth, Mukerji, and Mavko 2010) to ob-
+tain time-lapse subsurface models (Fig. 1(d)). We used this
+model because at higher pressure condition, local fluid flow
+slows down resulting in an acoustic velocity trend which fol-
+lows patchy saturation (Li et al. 2018b).
+Based on these models, we then simulated the baseline
+seismic survey corresponding to the initial stage (before the
+injection of CO 2) and the monitor seismic survey corre-
+sponding to the final stage at the end of the reservoir sim-
+ulation (Fig. 1(e)). As mentioned in the introduction, it is
+very difficult to exactly replicate the baseline and moni-
+tor surveys. In order to mimic the realistic scenario in the
+field, the baseline and monitor datasets were simulated us-
+ing different acquisition geometries (position of the mea-
+surements). Finally, we recovered the time-lapse seismic im-
+ages using JRM (Oghenekohwo and Herrmann 2017b; Wa-
+son, Oghenekohwo, and Herrmann 2017; Yin, Louboutin,
+and Herrmann 2021) to alleviate potential noise and inac-
+curacies in the seismic images in the case of non-replicatedtime-lapse surveys. These recovered images along with the
+label (leakage/no-leakage) serve as the input to the classi-
+fication network. We generated a total of 1000 leakage and
+870no-leakage scenarios, and computed the baseline, moni-
+tor and difference images with the JRM method in each case.
+Training
+The seismic difference images (difference between baseline
+and monitor recovery results) were converted to 224x224
+gray-scale images with bi-linear interpolation and trans-
+formed into three channel images where each channel is a
+copy of the actual gray-scale image. For the classification,
+the image dataset was randomly split into an 80% train-
+ing set and 20% test set. The training set was then further
+divided into two parts, one for model parameter updating
+(training) and another for hyperparameter tuning (valida-
+tion). The training hyperparameters from this second part
+are summarized in Table 1. For training, we replaced the last
+fully connected layers (classification layers) of each model
+with a new fully connected layer. We then trained the net-
+work (Fig. 1(g)) in two steps. First, we only trained the
+last classification layer, by freezing all the other layers, for
+100 epochs. Since most of the layers are fixed and do not
+need gradient updates, this first stage is extremely cheap and
+computationally efficient. Second, we further trained the full
+model for an additional 30 epochs to allow fine-tuning of all
+layers for our specific classification task. Following standard
+practices in classification settings, we used a binary cross-
+entropy loss function and the Adam optimizer (Kingma and
+Ba 2015) for all models. Finally, after the training (Fig. 1
+(h)), we implemented the CAM based methods (Fig. 1 (i)).
+We used the last convolutional layer activations for the CNN
+models, and the activations preceding the last attention layer
+for the transformer-based models.
+Analysis
+We show on Table 2, different performance metrics on our
+testing dataset, after training our four networks, with means
+and confidence intervals after 15 different runs. In detail, we
+show standard metrics such as accuracy, precision, and re-
+call. Additionally, we also show F1 score (Chinchor 1992),
+
+Model Accuracy Precision Recall F1 ROC-AUC
+VGG16 0.920 ±0.089 0.941 ±0.133 0.921 ±0.081 0.927 ±0.075 0.920 ±0.076
+ResNet34 0.948±0.020 0.982 ±0.028 0.928 ±0.044 0.948 ±0.040 0.967 ±0.019
+ViT 0.857 ±0.018 0.910 ±0.102 0.820 ±0.098 0.859 ±0.036 0.923 ±0.023
+Swin 0.836 ±0.036 0.881 ±0.108 0.818 ±0.078 0.841 ±0.076 0.909 ±0.007
+Table 2: Comparison of performance (for precision and recall, positives represent leakage whereas negatives are no leakage) on the test
+dataset for our four neural networks. The highest performance for each metric is highlighted in bold.
+Figure 2: Grad-CAM and Score-CAM saliency maps overlayed on the corresponding input seismic image containing a CO 2plume from
+leakage. The CO 2plume can be seen on the seismic image as the high amplitude event at 1.3km depth and 1.8km in X.
+that combine recall and accuracy, and area under curve of re-
+ceiver operating characteristic (ROC-AUC) (Bradley 1997)
+to further evaluate the classification performance of mod-
+els. We observe in Table 2 that the CNN models outperform
+the transformer variants in all the metrics by a significant
+margin and that ResNet34 achieves the best performance in
+all the measures of evaluation. This result is consistent with
+the literature, hinting that despite being very accurate on a
+specific task, transformers do not generalize well with our
+modest sized dataset (Dosovitskiy et al. 2021). Additionally,
+we observe that all models lead to better precision compared
+to recall (more false negatives than false positives). This dis-
+crepancy can be attributed to the fact that certain leakage
+images have very small CO 2leakage areas (up to a single
+pixel) in the seismic images and are consequently very diffi-
+cult to detect.
+Second, we show in Fig. 2 the CAM results of each model
+on a single seismic image from our test dataset. The high
+amplitude area shows the regions of the seismic images
+that are most important to the classifier. As expected, those
+heatmaps provide an explainable representation of the clas-
+sification as the high amplitudes align with the CO 2leakage
+part of the seismic image. We observe that for the CNN, the
+saliency maps are well centered on the CO 2leakage por-
+tion despite being very coarse. Because of this coarseness,
+both Grad-CAM and score-CAM provide similar results. On
+the other hand, transformer-based networks lead to more fo-cused saliency maps that target the location of the CO 2leak-
+age extremely well. We observe in that case, the Score-CAM
+leads to reduction of aliases and noise compared to the Grad-
+CAM results. This can be linked to the potential presence of
+noise in the gradients of the transformers as the networks are
+very deep (Wang et al. 2020).
+Conclusion
+We have introduced an interpretable deep-learning method
+for CO 2leakage detection with very high accuracy on a
+synthetic but realistic model of a CO 2sequestration reser-
+voir. First, we showed through four state-of-the-art models
+that we can detect potential CO 2leakage from the recov-
+ered time-lapse seismic images. Second, we demonstrated
+that CAM provides an interpretable and accurate visual-
+ization of the CO 2plume in case of leakage. Addition-
+ally, we showed that transformer-based models (ViT, Swin)
+led to more focused CAM and that Score-CAM provided
+cleaner and therefore more explainable heatmaps. On the
+other hand, we found that standard CNNs led to better classi-
+fication results and therefore better leakage detection. In par-
+ticular, ResNet model performed best and achieved a very
+high score above 90% in every evaluation metric. Future
+work will focus on improving the classification network to
+achieve higher accuracy in leakage detection and on refining
+the heatmaps for better explainability.
+
+Acknowledgments
+This research was carried out with the support of Georgia
+Research Alliance and partners of the ML4Seismic Center.
+The authors thank Philipp A. Witte at Microsoft for the con-
+structive discussion.
+References
+Arts, R. J.; Chadwick, A.; Eiken, O.; Thibeau, S.; and
+Nooner, S. L. 2008. Ten years’ experience of monitoring
+CO2 injection in the Utsira Sand at Sleipner, offshore Nor-
+way. First Break , 26.
+Avseth, P.; Mukerji, T.; and Mavko, G. 2010. Quantitative
+seismic interpretation: Applying rock physics tools to reduce
+interpretation risk . Cambridge university press.
+Ayeni, G.; and Biondi, B. 2010. Target-oriented joint least-
+squares migration/inversion of time-lapse seismic data sets.
+Geophysics , 75.
+Bradley, A. P. 1997. The use of the area under the ROC curve
+in the evaluation of machine learning algorithms. Pattern
+Recognition , 30(7): 1145–1159.
+Chadwick, A.; Williams, G.; Delepine, N.; Clochard, V .; La-
+bat, K.; Sturton, S.; Buddensiek, M.-L.; Dillen, M.; Nickel,
+M.; Lima, A. L.; Arts, R.; Neele, F.; and Rossi, G. 2010.
+Quantitative analysis of time-lapse seismic monitoring data
+at the Sleipner CO2 storage operation. The Leading Edge ,
+29(2): 170–177.
+Chinchor, N. 1992. MUC-4 Evaluation Metrics. In Pro-
+ceedings of the 4th Conference on Message Understanding ,
+MUC4 ’92, 22–29. USA: Association for Computational
+Linguistics. ISBN 1558602739.
+Costa, A. 2006. Permeability-porosity relationship: A reex-
+amination of the Kozeny-Carman equation based on a frac-
+tal pore-space geometry assumption. Geophysical Research
+Letters , 33(2).
+Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn,
+D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.;
+Heigold, G.; Gelly, S.; Uszkoreit, J.; and Houlsby, N. 2021.
+An Image is Worth 16x16 Words: Transformers for Image
+Recognition at Scale. ICLR .
+E. Jones, C.; A. Edgar, J.; I. Selvage, J.; and Crook, H. 2012.
+Building Complex Synthetic Models to Evaluate Acquisi-
+tion Geometries and Velocity Inversion Technologies. Eu-
+ropean Association of Geoscientists & Engineers , cp-293-
+00580.
+Furre, A.-K.; Eiken, O.; Alnes, H.; Vevatne, J. N.; and
+Kiær, A. F. 2017. 20 Years of Monitoring CO2-injection
+at Sleipner. Energy Procedia , 114: 3916–3926. 13th Inter-
+national Conference on Greenhouse Gas Control Technolo-
+gies, GHGT-13, 14-18 November 2016, Lausanne, Switzer-
+land.
+He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep Resid-
+ual Learning for Image Recognition. In Proceedings of the
+IEEE conference on computer vision and pattern recogni-
+tion,, 770–778.
+IPCC. 2018. Global warming of 1.5°C. An IPCC Special
+Report on the impacts of global warming of 1.5°C abovepre-industrial levels and related global greenhouse gas emis-
+sion pathways, in the context of strengthening the global re-
+sponse to the threat of climate change, sustainable develop-
+ment, and efforts to eradicate poverty. In Press .
+Kingma, D. P.; and Ba, J. 2015. Adam: A Method for
+Stochastic Optimization. arXiv preprint arXiv:1412.6980.
+Li, B.; Zhou, F.; Li, H.; Duguid, A.; Que, L.; Xue, Y .; and
+Tan, Y . 2018a. Prediction of CO2 leakage risk for wells in
+carbon sequestration fields with an optimal artificial neural
+network. International Journal of Greenhouse Gas Control ,
+68: 276–286.
+Li, D.; Wei, J.; Di, B.; Ding, P.; Huang, S.; and Shuai, D.
+2018b. Experimental study and theoretical interpretation of
+saturation effect on ultrasonic velocity in tight sandstones
+under different pressure conditions. Geophysical Journal
+International , 212: 2226–2237.
+Li, D.; Xu, K.; Harris, J. M.; and Darve, E. 2020. Cou-
+pled Time-Lapse Full-Waveform Inversion for Subsurface
+Flow Problems Using Intrusive Automatic Differentia-
+tion. Water Resources Research , 56(8): e2019WR027032.
+E2019WR027032 10.1029/2019WR027032.
+Liu, Z.; Lin, Y .; Cao, Y .; Hu, H.; Wei, Y .; Zhang, Z.; Lin, S.;
+and Guo, B. 2021. Swin Transformer: Hierarchical Vision
+Transformer using Shifted Windows. In ICCV .
+Lumley, D. E. 2001. Time-lapse seismic reservoir monitor-
+ing.GEOPHYSICS , 66(1): 50–53.
+Oghenekohwo, F.; and Herrmann, F. J. 2017a. Highly re-
+peatable time-lapse seismic with distributed Compressive
+Sensing–-mitigating effects of calibration errors. The Lead-
+ing Edge , 36(8): 688–694. (The Leading Edge).
+Oghenekohwo, F.; and Herrmann, F. J. 2017b. Improved
+time-lapse data repeatability with randomized sampling and
+distributed compressive sensing. In EAGE Annual Confer-
+ence Proceedings . (EAGE, Paris).
+Ringrose, P. 2020. How to store CO2 underground: Insights
+from early-mover CCS Projects , volume 129. Springer.
+Ringrose, P.; Mathieson, A.; Wright, I.; Selama, F.; Hansen,
+O.; Bissell, R.; Saoula, N.; and Midgley, J. 2013. The In
+Salah CO2 Storage Project: Lessons Learned and Knowl-
+edge Transfer. Energy Procedia , 37: 6226–6236. GHGT-11
+Proceedings of the 11th International Conference on Green-
+house Gas Control Technologies, 18-22 November 2012,
+Kyoto, Japan.
+Russakovsky, O.; Deng, J.; Su, H.; Krause, J.; Satheesh, S.;
+Ma, S.; Huang, Z.; Karpathy, A.; Khosla, A.; Bernstein, M.;
+et al. 2015. Imagenet large scale visual recognition chal-
+lenge. International journal of computer vision , 115(3):
+211–252.
+Selvaraju, R. R.; Cogswell, M.; Das, A.; Vedantam, R.;
+Parikh, D.; and Batra, D. 2019. Grad-CAM: Visual Explana-
+tions from Deep Networks via Gradient-Based Localization.
+International Journal of Computer Vision , 128(2): 336–359.
+Simonyan, K.; and Zisserman, A. 2014. Very Deep Convolu-
+tional Networks for Large-Scale Image Recognition. arXiv
+preprint arXiv:1409.1556 .
+
+Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones,
+L.; Gomez, A. N.; Kaiser, L.; and Polosukhin, I. 2017. At-
+tention Is All You Need. In Advances in Neural Information
+Processing Systems , 6000–6010.
+Wang, H.; Wang, Z.; Du, M.; Yang, F.; Zhang, Z.; Ding, S.;
+Mardziel, P.; and Hu, X. 2020. Score-CAM: Score-Weighted
+Visual Explanations for Convolutional Neural Networks. In
+CVPR .
+Wason, H.; Oghenekohwo, F.; and Herrmann, F. J. 2017.
+Low-cost time-lapse seismic with distributed compressive
+sensing–-Part 2: impact on repeatability. Geophysics , 82(3):
+P15–P30. (Geophysics).
+Wen, G.; Tang, M.; and Benson, S. M. 2021. Towards a
+predictor for CO2 plume migration using deep neural net-
+works. International Journal of Greenhouse Gas Control ,
+105: 103223.
+Witte, P. A.; Louboutin, M.; Kukreja, N.; Luporini, F.;
+Lange, M.; Gorman, G. J.; and Herrmann, F. J. 2019.
+A large-scale framework for symbolic implementations of
+seismic inversion algorithms in Julia. Geophysics , 84(3):
+F57–F71. (Geophysics).
+Yin, Z.; Louboutin, M.; and Herrmann, F. J. 2021. Compres-
+sive time-lapse seismic monitoring of carbon storage and se-
+questration with the joint recovery model. In SEG Technical
+Program Expanded Abstracts , 3434–3438. (IMAGE, Den-
+ver).
+Yosinski, J.; Clune, J.; Bengio, Y .; and Lipson, H. 2014.
+How transferable are features in deep neural networks?.
+In Advances in neural information processing systems ,
+3320–3328.
+Zhou, B.; Khosla, A.; Lapedriza, A.; Oliva, A.; and Torralba,
+A. 2015. Learning Deep Features for Discriminative Local-
+ization. In IEEE CVPR, , 2921–2929.
+Zhou, Z.; Lin, Y .; Zhang, Z.; Wu, Y .; Wang, Z.; Dilmore,
+R.; and Guthrie, G. 2019. A data-driven CO2 leakage de-
+tection using seismic data and spatial-temporal densely con-
+nected convolutional neural networks. International Journal
+of Greenhouse Gas Control , 90: 102790.
+

aaaifss2022_15.txt

@@ -0,0 +1,323 @@
+Predicting Wildfire Risk Under Novel 21st-Century Climate Conditions
+Matthew Cooper
+Sust Global
+595 Pacific Ave., Floor 4
+San Francisco, California 94133
+Abstract
+Wildfires are one of the most impactful hazards associ-
+ated with climate change, and in a hotter, drier world,
+wildfires will be much more common than they have
+historically been. However, the exact severity and fre-
+quency of future wildfires are difficult to estimate, be-
+cause climate change will create novel combinations of
+vegetation and fire weather outside what has been his-
+torically observed. This provides a challenge for AI-
+based approaches to long-term fire risk modeling, as
+much future fire risk is outside of the available feature
+space provided by the historical record. Here, I give an
+overview of this problem that is inherent to many cli-
+mate change impacts and propose a restricted model
+form that makes monotonic and interpretable predic-
+tions in novel fire weather environments. I then show
+how my model outperforms other neural networks and
+logistic regression models when making predictions on
+unseen data from a decade into the future.
+Introduction
+One way to describe the future effects of climate change is
+with the phrase global weirding . The 21st century will be in-
+creasingly uncanny, as we will see Caribbean beach weather
+in Iceland; deserts that become soggy and green; and an Arc-
+tic Ocean that is entirely free of ice, potentially by 2035
+(Guarino et al. 2020). Novel assemblages of temperature,
+precipitation, land cover, and vegetation will emerge that are
+unlike anything in human history, giving rise to hazards un-
+precedented in severity and posing major challenges to adap-
+tation. Additionally, these weird conditions are a challenge
+to any form of modeling that depends on rich training data,
+as much of the future will be entirely outside of the feature
+space of available observational data.
+This is especially true in the case of wildfire, because fire
+depends on two things: burnable vegetation and dry enough
+conditions to ignite that vegetation. Under stable climate
+conditions, weather and vegetation reach an equilibrium,
+where the amount of burnable vegetation is proportional to
+the amount of rainfall (See Fig. 1). However, under climate
+change, we are seeing increasingly novel pairings of pre-
+cipitation and vegetation (See Fig. 2). For example, Califor-
+Copyright © 2022, Association for the Advancement of Artificial
+Intelligence (www.aaai.org). All rights reserved.nia has historically had dry summers and wet winters, lead-
+ing to chaparral and spare forest vegetation communities.
+However, in the past decade, California had weather condi-
+tions more characteristic of a desert climate. This extremely
+dry weather, coupled with high levels of vegetation, is what
+has caused the unprecedented fire crisis in California (Abat-
+zoglou and Williams 2016). A similar situation is occurring
+in the Amazon, where tropical rainforest vegetation is expe-
+riencing increasingly long dry seasons and is converting into
+a tropical savanna, with fire consuming the excess biomass
+(Le Roux et al. 2022).
+These emerging conditions are causing significant prob-
+lems for sectors like the insurance industry, which has
+traditionally used historic risk to estimate future risk and
+appropriately price premiums. Unable to accurately esti-
+mate fire risk under unprecedented conditions, many home
+insurance companies are withdrawing from fire-prone ar-
+eas, leaving homeowners without coverage (Poizner 2022;
+Singh 2022). Given that a typical home mortgage can last
+up to 30 years, a period over which climatological and eco-
+logical systems will continue to disequilibrate, it is impera-
+tive that we develop better methods for estimating fire risk
+that can make reasonable predictions outside of the existing
+feature space provided by historic data.
+Data
+For this analysis, I use data on fire occurrence provided glob-
+ally and at a 500 meter resolution derived from NASA’s
+MODIS satellite program (Giglio et al. 2009). This dataset
+goes back to November 2000 and provides a binary indica-
+tor of whether a fire was observed at a given pixel at a daily
+timestep. From this dataset, I collected 240 million sample
+locations on a given day across the terrestrial world, over-
+sampling fire occurrence to make up approximately 10% of
+the dataset, but otherwise sampling completely at random.
+For each sample point, I calculate a daily fire weather in-
+dex known as the Keetch-Byram Drought Index, or KBDI
+(Brown, Wang, and Feng 2021; Gannon and Steinberg
+2021). KBDI is an index updated on a daily time step and is
+indicative of the amount of water in the top 203 millimeters
+of soil. A KBDI score of 0 corresponds to saturated soil and
+very little fire risk, while a KBDI score of 203 indicates that
+soil is dry up to 203 millimeters deep and that fire risk is very
+high. To calculate historic values of this index, I use daily
+
+Figure 1: Historically, precipitation and biomass have been
+in equilibrium.
+HIGH Biomass 
+Rainfall HIGH LOW LOW 
+Figure 2: Under climate change, precipitation and biomass
+are decoupled, leading to unprecedented fire severity in Cal-
+ifornia and the Amazon.
+HIGH Biomass Rainfall HIGH LOW 
+LOW 
+Amazon 
+Wildfires California 
+Wildfires 
+historic data on temperature and precipitation from the 10
+kilometer ERA5-Land reanalysis dataset (Mu ˜noz-Sabater et
+al. 2021). Additionally, to better determine the fire risk con-
+text I determine the local climate zone for each point using
+the Koppen-Geiger methodology (K ¨oppen 2011), as well as
+the local land cover type using the 300 meter ESA land cover
+dataset (ESA 2017).
+For my analysis, I use observed data from November 2000
+to October 2011 as my training data ( n= 135,559), and ob-
+served data from November 2011 to October 2021 as my
+validation data ( n= 123,428). Testing my model on obser-
+vations that occurred a decade beyond the end of the train-
+ing data can give me an indication of how my model will
+perform over the course of the next decade. Additionally, I
+subset my analysis to eastern Oregon to constrain the discus-
+sion, although I have data processed and prepared for analy-
+ses at a global scale.
+Finally, for future estimates of fire weather to use a fea-
+tures in model inference, I derive KBDI from ensembled
+and bias-corrected simulations of temperature and precipita-
+tion throughout the 21st century using Global Climate Mod-
+els (GCMs) from the 6th Climate Model Intercomparison
+Project (CMIP6) (O’Neill et al. 2016).The Problem
+To better illustrate the modeling challenge presented by
+novel fire conditions, also referred to as domain shift, I show
+daily fire weather values (KBDI) in eastern Oregon for peri-
+ods where observed KBDI scores were indicative of elevated
+fire risk (KBDI >100), typically in the summer (See Fig.
+3). Eastern Oregon is an area without significant historic fire
+activity but is increasingly threatened by fire. There, KBDI
+values are increasing every decade, with the next decade
+modeled to have KBDI values at the maximum potential fire
+risk. This prevalence of increasingly out-of-sample and un-
+precedented fire weather is also associated with heightened
+fire risk, something models trained on only historic data will
+struggle to capture.
+Figure 3: Shifting of fire weather towards unprecedented risk
+each decade complicates empirical AI modeling. Histogram
+of daily KBDI values in Eastern Oregon, by decade. Values
+for 2000-2010 and 2011-2021 are observed, values for 2022-
+2032 are taken from an ensemble of bias-corrected climate
+models.
+I further illustrate this domain shift modeling challenge
+by training a simple 3-layer feed-forward neural network to
+predict the probability of fire in eastern Oregon as a function
+of KBDI using sample data from 2000-2011 and validation
+data from 2012-2022. I compare that model against a logis-
+tic regression model using the same dataset. I find that the
+neural network under-estimated fire risk at high KBDI lev-
+els, while the logistic regression, due to its implicit mono-
+tonicity, better captured the trend of increasing fire risk with
+increasing KBDI levels (See Fig. 4).
+While these test datasets illustrate the nature of the prob-
+lem, both models used here were quite simple. In addition
+to fire weather, fire risk is heavily determined by other con-
+textual factors, including biomass, land cover, long-term cli-
+mate conditions, and elevation. I therefore construct more
+complex models based on 24 features derived from my sam-
+ple dataset, one-hot encoding for land cover type and climate
+zone, as well as including terms for latitude and longitude,
+allowing the models to learn location-specific fire risk rela-
+tionships. Additionally, I fit a hierarchical logistic regression
+using the same features as the multivariate neural network.
+Overall, I find that multivariate models perform better
+than univariate models based only on KBDI when evaluated
+on a held out test dataset from the next decade (See Table
+1). Additionally, I find that logistic regression models out-
+perform neural networks on the test data, because they make
+
+Figure 4: Observed probability of fire by KBDI value, in the
+training and testing datasets. Additionally, I show the predic-
+tions of a simple feed-forward neural network and a logis-
+tic regression. Note that the neural network under-estimates
+out-of-sample future fire risk.
+predictions that are monotonic. This suggests that the neural
+networks struggle to capture extreme behavior.
+New Architecture
+Because simple neural networks struggle to capture fire ex-
+tremes under novel data domains, I propose a new neural
+network architecture, based on two premises. The first is
+that the relationship between KBDI and fire probability is
+monotonic, and as ongoing climate change leads to condi-
+tions drier than any previously observed in many locations,
+it will be necessary to use models that can extrapolate mono-
+tonically, such as logistic regression models. Secondly, the
+parameterization of the weather-fire relationship is complex
+and context dependent, with a large number of influenc-
+ing variables that interact nonlinearly, requiring models like
+neural networks that can handle such estimation problems.
+Drawing from both of these premises, I have implemented
+a neural network architecture that uses a large number of
+features describing the geographic context to estimate the
+parameters of a logistic model that describes the KBDI-fire
+relationship in that context. In this case, I use features for
+the spatial location, local land cover type, and historic cli-
+mate zones indicative of prevailing vegetation communities;
+however, this architecture could be extended to incorporate
+other important features, such as topography, proximity to
+human settlements, or aboveground biomass. This approach
+has the advantage of drawing on complex interactions within
+the geophysical environment that influence the relationship
+between fire and weather conditions, while still being con-
+strained to make predictions in line with my strong prior as-
+sumption that the relationship between dryness and fire risk
+is monotonic.
+The model feeds a large number of features in four dense
+hidden layers that condense from 32 to 8 nodes with a ReLU
+activation function. The model then diverges into two sepa-
+rate hidden layers, each of which converges into a single-
+parameter output, which are treated as the two parameters
+in a logistic regression ( 0and1). The model’s loss func-
+tion is therefore the performance of those two parameters in
+a logistic regression using observed KBDI, evaluated with
+binary cross-entropy (See Fig. 5).Figure 5: Diagrammatic representation of fire neural net-
+work used to estimate logistic regression parameters.
+Linear 
+Predictor 
+β1 β0y
+Binarized 
+Cross- 
+Entropy (       , ) β0 β1Linear 
+Pred. y +24 Input Features 
+X 
+8-Node Dense 8-Node Dense 16-Node Dense 32-Node Dense 
+8-Node Dense 8-Node Dense 
+Loss Function 
+Model R2MSE
+Univariate NN 0.0091 0.0442
+Logistic Regression 0.0139 0.0440
+Multivariate NN 0.0156 0.0439
+Hierarchical Logistic Regression 0.0166 0.0438
+NN-Estimated Logistic Regression 0.0202 0.0436
+Table 1: Model performance by R2and mean squared error
+(MSE).
+I fit a model with this architecture using the same fea-
+tures as the aforementioned multivariate neural network and
+find that it improves performance on R2by 22%. This archi-
+tecture is able to draw on the advantages of using gradient
+descent to explore complex relationships among features,
+while still making predictions that are interpretable and ex-
+trapolate well outside of the observed range of fire weather
+values.
+Conclusion
+While there would be many benefits of using this method-
+ology, it would have the drawback of requiring a very large
+dataset, as is typical of neural network based approaches.
+This would evolve the state of the art of predicting wildfires
+by focusing specifically on making predictions outside of
+the feature space available for training. Having better long-
+
+term fire predictions would help state agencies and govern-
+ments to eliminate risks, as they currently rely on projec-
+tions that are more near-term, focusing on weekly to sea-
+sonal timescales.
+Neural networks provide a number of advantages and can
+explore a hyper-dimensional and complex feature space ef-
+ficiently. However, they are brittle outside of their training
+space. In such situations where it is necessary to make pre-
+dictions in the absence of available training data, predictions
+must be guided by theory and model behavior must be in-
+terpretable. I therefore developed an architecture that flex-
+ibly draws on complex environmental variables while still
+making predictions that are aligned with my theoretical prior
+that drier weather leads to increased fire risk. I find that this
+model performs better than other approaches when used to
+make predictions a decade into the future. Given the theoret-
+ical support of this approach, it is likely to be especially use-
+ful for making estimates at even longer timescales of up to
+two or three decades. This approach has relevance for mod-
+eling many of the novel risks posed by climate change.
+References
+Abatzoglou, J. T., and Williams, A. P. 2016. Impact of an-
+thropogenic climate change on wildfire across western US
+forests. Proc. Natl. Acad. Sci. U.S.A. 113(42):11770–11775.
+Brown, E. K.; Wang, J.; and Feng, Y . 2021. US wildfire
+potential: a historical view and future projection using high-
+resolution climate data. Environ. Res. Lett. 16(3):034060.
+ESA. 2017. Land cover cci product user guide. Technical
+report.
+Gannon, C. S., and Steinberg, N. C. 2021. A global as-
+sessment of wildfire potential under climate change utilizing
+keetch-byram drought index and land cover classifications.
+Environmental Research Communications 3(3):035002.
+Giglio, L.; Loboda, T.; Roy, D. P.; Quayle, B.; and Justice,
+C. O. 2009. An active-fire based burned area mapping
+algorithm for the MODIS sensor. Remote Sens. Environ.
+113(2):408–420.
+Guarino, M.-V .; Sime, L. C.; Schr ¨oeder, D.; Malmierca-
+Vallet, I.; Rosenblum, E.; Ringer, M.; Ridley, J.; Feltham,
+D.; Bitz, C.; Steig, E. J.; et al. 2020. Sea-ice-free arctic
+during the last interglacial supports fast future loss. Nature
+Climate Change 10(10):928–932.
+K¨oppen, W. 2011. The thermal zones of the earth accord-
+ing to the duration of hot, moderate and cold periods and to
+the impact of heat on the organic world. Meteorologische
+Zeitschrift 20(3):351–360.
+Le Roux, R.; Wagner, F.; Blanc, L.; Betbeder, J.; Gond, V .;
+Dessard, H.; Funatzu, B.; Bourgoin, C.; Cornu, G.; Herault,
+B.; Montfort, F.; Sist, P.; Begue, A.; Dubreuil, V .; Laurent,
+F.; Messner, F.; Hasan, A. F.; and Arvor, D. 2022. How
+wildfires increase sensitivity of Amazon forests to droughts.
+Environ. Res. Lett. 17(4):044031.
+Mu˜noz-Sabater, J.; Dutra, E.; Agust ´ı-Panareda, A.; Al-
+bergel, C.; Arduini, G.; Balsamo, G.; Boussetta, S.; Choulga,
+M.; Harrigan, S.; Hersbach, H.; Martens, B.; Miralles, D. G.;Piles, M.; Rodr ´ıguez-Fern ´andez, N. J.; Zsoter, E.; Buon-
+tempo, C.; and Th ´epaut, J.-N. 2021. ERA5-Land: a state-of-
+the-art global reanalysis dataset for land applications. Earth
+Syst. Sci. Data 13(9):4349–4383.
+O’Neill, B. C.; Tebaldi, C.; Van Vuuren, D. P.; Eyring, V .;
+Friedlingstein, P.; Hurtt, G.; Knutti, R.; Kriegler, E.; Lamar-
+que, J.-F.; Lowe, J.; et al. 2016. The scenario model inter-
+comparison project (scenariomip) for cmip6. Geoscientific
+Model Development 9(9):3461–3482.
+Poizner, S. 2022. Op-Ed: Wildfires never threatened my
+home. But my insurer said they do — and dumped me. Los
+Angeles Times .
+Singh, A. G. 2022. The need to modernize california wild-
+fire insurance regulation with climate science. Journal of
+Science Policy and Governance 20(1).
+

aaaifss2022_16.txt

@@ -0,0 +1,1070 @@
+Probabilistic Machine Learning in Polar Earth and Climate Science:
+A Review of Applications and Opportunities
+Kim Bente,1Judy Kay,1Roman Marchant2
+1School of Computer Science, The University of Sydney, Australia
+2CSIRO’s Data61, Australia
+[email protected], [email protected], [email protected]
+Abstract
+Our world’s climate future is on thin ice. The study of long-
+term weather patterns in the polar regions is an important
+building block in tackling Climate Change. Our understand-
+ing of the past, the present and the future of the earth sys-
+tem, and the inherent uncertainty, informs planning, mitiga-
+tion, and adaptation strategies. In this work we review pre-
+vious applications of machine learning and statistical com-
+puting to polar climate research, and we highlight promising
+probabilistic machine learning methods that address the mod-
+elling needs of climate-related research in the Arctic and the
+Antarctic. We discuss common challenges in this interdisci-
+plinary field and provide an overview of opportunities for fu-
+ture work in this novel area of research.
+Introduction and Background
+This section introduces and defines the Polar Earth and
+Climate Science domain and substantiates the urgent need
+for continuing research in this field, most importantly to
+inform policy and decision-making based on a scientific,
+uncertainty-aware foundation. Next, we give a brief back-
+ground on the recent growth in machine learning (ML) to
+address Climate Change, propelled by an increase in data
+availability, the simultaneous leaps in computing power, and
+advances in artificial intelligence (AI) and machine learning
+methods. We motivate the emphasis on using a probabilistic
+framework to convey uncertainty modelling needs.
+With this uncertainty-aware perspective, this paper con-
+tributes a review of important machine learning applications
+in the polar parts of the Earth and Climate Science domain,
+thus building on the wider work of Rolnick et al. (2019).
+We discuss methodological aspects and the types of do-
+main problems addressed in previous work to then synthe-
+sise common challenges. We introduce suitable probabilistic
+machine learning methods, particularly Bayesian Optimisa-
+tion and causal methods, and highlight novel research from
+these areas where we recognise strong opportunities for fu-
+ture work in polar climate applications.
+Polar Earth and Climate Science
+Climate Change is one of the greatest challenges humanity is
+facing today. While on average, our globe is warming, tem-
+Copyright © 2022, Association for the Advancement of Artificial
+Intelligence (www.aaai.org). All rights reserved.peratures in the Arctic have increased by more than double
+the global mean over the last two decades (IPCC 2019). Pro-
+jections from a framework of state-of-the-art physics-based
+climate simulation models, CMIP6 (the acronym for Cou-
+pled Model Intercomparison Project phase 6) (Eyring et al.
+2016), predict that the Arctic ocean will become largely free
+of sea ice during summer months by 2050, even under op-
+timistic anthropogenic emission scenarios (Notz and Com-
+munity 2020). Contributions from ice sheets and glaciers,
+notably the Greenland Ice sheet and the West Antarctic Ice
+Sheet, are understood to be the dominant source of the rise in
+sea level (IPCC 2019). This poses a direct threat to the liveli-
+hoods of a large number of people who live on low-lying is-
+lands, in coastal regions but also in inland, flood prone areas.
+These and other concerning changes like ocean acidifica-
+tion resulting from absorption of anthropogenic CO2 emis-
+sions, ocean warming (IPCC 2019), or the acceleration in
+the Antarctic Circumpolar Current (Shi et al. 2021) highlight
+the critical role the polar regions hold in the context of the
+climate system: The cryosphere, describing all frozen water
+part of the Earth system, as well as the neighboring oceans,
+are strongly linked to other components of the global climate
+system through the exchange of carbon, water and energy
+(IPCC 2019).
+Accelerated by the recent pace of change and the loom-
+ing threats to livelihoods and ecosystems, there are strong
+academic efforts in further growing our understanding of
+the field. Earth Science and Climate Science are both well-
+established research areas. The Earth Sciences are tradition-
+ally decomposed into the five interacting systems of earth,
+namely the atmosphere, the hydrosphere, the biosphere, the
+geosphere, and the cryosphere. Climate Science is the study
+of long-term weather patterns, primarily investigating atmo-
+spheric properties, but also building on the other subsystems
+of the Earth Sciences by studying interactions with, for ex-
+ample, the ocean, or, over longer timescales, the geosphere
+(Springer Nature 2022). To illustrate the interconnection of
+these disciplines, ice cores from Antarctica for instance, en-
+able paleoclimatology researchers to determine past concen-
+trations of greenhouse gases in the atmosphere. To do so,
+they analyse air bubbles which were trapped in the ice up to
+a million years ago. Thus, discoveries in the Earth Sciences
+often seed new insights for Climate. The problem of map-
+ping the bedrock topography of Antarctica further show-
+
+Figure 1: Schematic diagram of selected aspects relevant to the intersection of the Probabilistic Machine Learning field and the
+Polar Earth and Climate Science domain discussed in this paper.
+cases how these fields are intertwined. This geology and
+Earth Sciences problem is directly related to the estimation
+of the ice topography and ice mass - quantities climate sci-
+entists are highly concerned with (Lythe and Vaughan 2001;
+Fretwell et al. 2013). Both examples fall under the umbrella
+of the Polar Sciences, a term that generally denotes scien-
+tific research from different disciplines relating to the polar
+regions (Elsevier 2022).
+Because of the urgency imposed by the rapidly chang-
+ing climate and its transnational scope, global organisations
+have formed and governments have committed to direct re-
+search resources, investments, and policy changes at this
+pressing issue. CSIRO, Australia’s national science agency,
+identified adapting to climate change as a global megatrend,
+with particular concern about natural disasters, climate-
+driven migration and impacts on water quality, infrastruc-
+ture and also public health (Naughtin et al. 2022). The In-
+tergovernmental Panel on Climate Change (IPCC), a body
+of the United Nations and paramount international platform,
+was created to assess the scientific foundations of Climate
+Change and to inform policy makers about their findings.
+Leading researchers from the various interconnected fields,
+including those studying earth and climate, contribute to
+the IPCC assessment reports. The most recent, sixth assess-
+ment cycle includes a ’Special Report on the Ocean and
+Cryosphere in a Changing Climate’ (IPCC 2019), empha-
+sising the need to deepen understanding in this dedicated
+domain.
+Machine Learning for Climate Research
+Whilst Earth and Climate Science are well-established ar-
+eas of research, and specifically research concerning the
+Arctic and Antarctic is advancing, the intersection of Arti-
+ficial Intelligence/Machine Learning/Data Science and the
+Climate Sciences is a fairly novel field. Within this com-
+munity the first of the annual Climate Informatics confer-
+ence series, referred to as Climate Informatics, was held in
+2011. Monteleoni, Schmidt, and McQuade (2013) provide
+an overview of new opportunities in this field, and in 2022
+a new journal, named Environmental Data Science, posi-tioned at the interface of Data Science and the environment,
+was established by leaders from the Climate Informatics
+community. Figure 1 presents a schematic overview of the
+two intersecting research areas and highlighted concepts re-
+viewed in this paper. These concepts, discussed throughout
+this section, are signaled with bold font. The data-focused
+and Earth Science research communities differ in their mod-
+elling paradigms, publishing norms, and research priorities.
+Despite these research silos, a growing community of re-
+searchers has acted on the great opportunity of truly interdis-
+ciplinary research and has established research organisations
+with the aim of combining powerful machine learning and
+statistical methods with the deep understanding of climate
+and earth system processes and the high-impact questions
+driving research in related fields. The organisation, Climate
+Change AI, emerged in 2019 from a series of workshops on
+‘Tackling Climate Change with Machine Learning’ at lead-
+ing machine learning conferences, as well as side events at
+the 2019 and 2021 United Nations Climate Change Confer-
+ences (COP25 and COP26 respectively) (Climate Change
+AI 2022). The eponymous paper (Rolnick et al. 2019) gives
+a big picture overview of problems associated with climate
+change where machine learning can be applied with impact.
+Rolnick et al. (2019) allocate areas of machine learning to
+suitable climate change solution domains, spanning mitiga-
+tion and adaptation strategies. Within climate prediction to
+inform adaptation strategies, Rolnick et al. outline impor-
+tant sub problems like data assimilation or the incorpora-
+tion of ice sheet dynamics into climate models to improve
+projections. In this paper, we aim to build on this overview,
+by deepening the review of machine learning applications to
+Polar Climate and Earth Science problems, and by outlining
+opportunities suitable to this specific domain.
+Remote sensing measurements (see Figure 1) from satel-
+lites and aircrafts, data from fixed monitoring stations,
+and field measurements from ice cores, roaming UA Vs, or
+oceanographic research vessels and floats (Shi et al. 2021)
+are all contributing to an increase in earth observation data
+available today. Continuous earth observations by satellite
+only started with Landsat 1 in 1972 (NASA 2021) so it can-
+
+not support the study of long-term climate patterns. Fortu-
+nately, indirect measurements of ice cores, rocks and corals,
+can provide data that goes multiple glacial periods back. Ad-
+vances in remote sensing technology allow a wide variety
+of properties to be directly measured or inferred, includ-
+ing altimetry, seismic activity, gravimetry, surface albedo,
+sea surface wind speeds or atmospheric properties. Data in
+this domain commonly have spatial and temporal dimen-
+sions (see Figure 1) and thus exhibit varying resolutions.
+These special characteristics can incur challenges with the
+data fusion and modelling process. Shirmard et al. (2022)
+provide a review of how machine learning and specifically
+deep learning is utilised to process various remote sensing
+data for mapping geological features - a use cases which is
+closely related to Climate Science applications. Overall, the
+data surge is a momentous opportunity to increase our un-
+derstanding of the least explored and less understood parts of
+the Earth, such as the oceans, the Arctic and Antarctic. To-
+gether with the simultaneous increase in computing power
+(hardware and algorithms) and the rise of machine learning
+and statistical computing, particularly in deep learning and
+causal inference methods (see Figure 1), this is creating vast
+opportunities to harness data-centric methods for scientific
+discovery.
+Uncertainty is an essential aspect of climate change data
+and its analysis. Predictions from climate models, together
+with their associated uncertainty, need to be interpreted to
+inform sensible decision-making. The uncertainty materi-
+alised in predictions arises from multiple sources and can
+be classified into measurement and model related. Some of
+these source of uncertainty arise from: physical limitations
+on sensors that place an upper bound on accuracy, data sets
+which can present biases, models with limited complexity
+which are imperfect representations of natural phenomena
+and inaccurate assumptions. We therefore believe that quan-
+tifying model uncertainty with probabilistic machine learn-
+ing methods, is important, especially in this domain. Many
+probabilistic machine learning methods are rooted in the
+Bayesian framework (see Figure 1), where model param-
+eters are represented with random variables, whose proba-
+bility distributions are used as a central tool to represent
+uncertainty on different layers of abstraction in the model.
+Furthermore, a fully Bayesian approach incorporates do-
+main expert knowledge through prior distributions, which
+after careful elicitation are combined with data and model
+assumptions to provide logically consistent and uncertainty
+aware estimations. We will therefore emphasise the perspec-
+tive of uncertainty quantification throughout this paper.
+Review of applications
+This section reviews machine learning and statistical com-
+puting applications for Polar Climate and Earth Science.
+Applications are grouped into climate model emulators, sea
+level rise prediction, topography mapping, sea ice forecast-
+ing, and lastly climate feedbacks and teleconnection. Table
+1 gives an overview of selected applications discussed, the
+geographical region addressed, the methods used, and the
+high-level discipline of the publishing venue.Climate model emulators
+State-of-the-art climate models, also known as Earth System
+Models (ESMs), simulate the interactions between the main
+climate drivers (atmosphere, land, ocean and ice) through
+physics-based coupled dynamics, to study the processes
+based on simulated data and to make predictions about fu-
+ture climate (Rolnick et al. 2019; Balaji et al. 2017). The
+latest state-of-the-art CMIP model, CMIP6 (Eyring et al.
+2016), is highly computationally expensive and data inten-
+sive (Balaji et al. 2017). This complexity arises because the
+model simulates a large set of different processes and sub-
+processes within and between the climate drivers, which
+take place on different time and spatial scales. Furthermore,
+CMIP6 is a multi-model ensemble of around 100 mod-
+els which were developed by over 50 different modelling
+groups (Copernicus 2021), scaling computational demands.
+One weakness of climate models is their sensitivity to small
+changes in initial conditions or other inputs (Balaji et al.
+2017), known as the butterfly effect from early chaos theory
+literature (Abraham and Ueda 2000). The characteristics of
+this challenge, i.e. to learn complex and often spatially dis-
+tant interactions within an uncertain environment, matches
+the potential of machine learning which can help with model
+estimation from fusing large amounts of multi-modal and
+disparate sources of data.
+To combat the computational and robustness issues of cli-
+mate models, deep learning can be used to create emulation
+models, which do not sacrifice accuracy but are computa-
+tionally highly efficient once trained (Reichstein et al. 2019).
+While climate models remain the benchmark for most gen-
+eral climate prediction tasks today, the use of machine learn-
+ing models to replace, complement or improve traditional
+first principle models is gaining momentum: Reichstein et al.
+propose to combine the strengths of theory-driven and data-
+driven modelling in a hybrid approach. Physical models are
+usually interpretable and deeply rooted in theoretical under-
+standing of the phenomenon, while machine learning mod-
+els are highly flexible and can adapt to data. Based on these
+different strengths of either paradigm, Reichstein et al. sug-
+gest that suitable domain problems replace physical sub-
+model components which are less well described by phys-
+ical theory, with machine learning models, which may even
+be able to learn unexpected patterns unknown to experts. Be-
+cause the cryosphere is a component of the earth system that
+is challenging to simulate (Gagn ´e, Gillett, and Fyfe 2015),
+this could be a great opportunity to apply deep learning emu-
+lation models. The authors of (Reichstein et al. 2019) further
+identify that machine learning models could also be used as
+a calibration layer on top of traditional models, to correct
+error patterns of the model. In addition Reichstein et al. em-
+phasise the need to quantify models’ credibility and confi-
+dence, specifically in the case of extrapolation. This could be
+achieved by using Bayesian Deep Learning Models, which
+bridge exactly this gap within deep learning (Chandra, Az-
+izi, and Cripps 2017). On a meta-level, decreasing the com-
+putational load for climate modelling will both speed up the
+process, and benefit the footprint of research in this field.
+
+Application Region Method Reference Venue category
+Emulation of climate
+modelsGlobal Deep Learning Reichstein et al. (2019) Interdisciplinary
+Sea level rise
+predictionAntarctic Hybrid probabilistic modelling
+[Statistics]Kopp et al. (2017) Earth & Climate
+Sea level rise
+predictionAntarctic Bayesian Hierarchical Models
+[Statistics]Zammit-Mangion et al.
+(2014, 2015)ML & Statistics
+Bedrock and ice
+topography mappingAntarctic Convolutional Neural Networks
+(CNN) [Deep Learning]Leong and Horgan
+(2020)Earth & Climate
+Sub-seasonal sea ice
+forecastingArctic Attention-based Ensemble Model
+(EA-LSTM) [Deep Learning]Ali et al. (2022) ML & Statistics
+Seasonal sea ice
+forecastingArctic U-Nets [Deep Learning] Andersson et al. (2021) Interdisciplinary
+Determining causal
+climate driversArctic Causal Effect Networks
+(CEN)[Causal Inference]Kretschmer et al. (2016) Earth & Climate
+Determining causal
+climate feedbacksAntarctic Convergent cross-mapping
+(CCM) [Causal Inference]van Nes et al. (2015) Earth & Climate
+Table 1: Overview of selected applications of machine learning (ML) and statistical computing methods to problems from the
+Polar Earth and Climate Science domain. The ’Venue category’ reflects the broad research community and is based on the
+subject area of the journal which the cited work is published in.
+Sea level rise predictions
+The prediction of sea level rise is an important problem due
+to its far reaching implications on human habitat. Because
+the mass balance (the sum of ice losses and gains) from
+the Greenland ice sheet, the Antarctic ice sheet and glaciers
+are the primary drivers of sea level rise (IPCC 2019), these
+modelling tasks are directly related to each other and con-
+sequently also to the dynamics of climate models (Rolnick
+et al. 2019). Government agencies like the United States’
+NOAA, Australia’s CSIRO, dedicated research groups like
+the Sea Level Research Group from CIRES at the Univer-
+sity of Colorado Boulder, or IMBIE, an international collab-
+oration of scientist led by the University of Leeds, all work
+in this field. The emission sensitivity in the predictions of
+the IPCC (2019) for mass loss is eminent. Especially in the
+high-emission scenario the accumulating uncertainty in pre-
+dicted global mean sea level rise is visible through the wide
+range of predicted increase at low confidence. In addition,
+sea level rise is not distributed uniformally around the globe
+(IPCC 2019). Particularly the modelling of ice loss in the
+Antarctic is recognised to be challenging. A recent mech-
+anistic understanding of accelerating effects from ice-shelf
+hydro-fracturing and collapsing of ice cliffs on mass loss,
+produces non-linear trends that far exceed established pre-
+dictions (Kopp et al. 2017). In this work Kopp et al. incor-
+porate an ensemble of Antarctic ice-sheet (AIS) simulations
+with a probabilistic framework. Kopp et al. argue strongly
+for the use of fully Bayesian models, and recommend for fu-
+ture work to identify domain-imposed constraints and well-
+informed prior beliefs over parameters.
+Aligning with the emphasis on probabilistic methodsto address this highly uncertain task, is the work of
+Gopalan, Zammit-Mangion, and McCormack. This pre-
+dicts the Antarctica’s contribution to sea-level rise using
+aBayesian Hierarchical Model (Zammit-Mangion et al.
+2014, 2015). On a high level, the different hierarchical lay-
+ers constitute of the parameter model, the process model
+(modelling latent dynamical processes), and the observa-
+tion model (Gopalan, Zammit-Mangion, and McCormack
+2021). Altimetry, gravimetry and GPS observations are
+used. Knowledge about multiple relevant physical processes
+is incorporated into the statistical model as prior distribu-
+tions and dependence structures, informed by traditional nu-
+merical ice dynamics models. A strong advantage of this
+technique is that all estimated quantities, not just predicted
+sea-level rise, have an associated credible interval reflecting
+uncertainty. Estimates, e.g. gravimetry parameter estimates,
+can be interpreted, offering insights for domain experts. Fur-
+ther, Gopalan, Zammit-Mangion, and McCormack (2021)
+used approximation methods to improve computational ef-
+ficiency. They provide an overview of Bayesian modelling
+and inference in glaciology, showcasing two projects, one
+being the above work by Zammit-Mangion et al. (2014).
+Topography mapping
+An understanding of the topography underneath the ice
+forms the basis for ice sheet modelling. The series of
+BedMap models, BedMap and the updated BedMap2, com-
+prise of gridded digital topographical models of the surface
+elevation, subglacial bed rock elevation, sea floor elevation,
+and also ice thickness for the continent of Antarctica (Lythe
+and Vaughan 2001; Fretwell et al. 2013). Data from various
+
+surveys, at different spatial scales, were assimilated to con-
+struct state-of-the-art mappings. The BedMap2 data set lays
+the foundation for many other researchers in this field. The
+dependence on up-stream estimates of quantities like sub-
+glacial bed rock elevation, which can not be directly mea-
+sured, exemplify the role of uncertainty within polar re-
+search. Building on top of BedMap2, Leong and Horgan
+(2020) introduce DeepBedMap to address the problem of
+imputing high spatial resolution bed elevation grids for ar-
+eas in Antarctica where no data at high resolution is avail-
+able. A variant of Deep Convolutional Neural Networks,
+adapted from Enhanced Super-Resolution Generative Ad-
+versarial Network, is used to generate high-resolution maps.
+Additional gridded data on ice surface elevation, velocity
+and snow accumulation, all available at high spatial resolu-
+tions, are used as inputs. To capture the spatial interaction of
+the different properties, the neural network was trained on
+ground truth data. Resulting surface roughness was evalu-
+ated as an indicator for realistic topography maps. Other re-
+cent work uses topographic satellite data to map supraglacial
+lakes in regions of Antarctica using Random Forest classi-
+fiers (Dirscherl et al. 2020). Despite the black-box character
+of such models, this showcases how machine learning can
+be used for assimilation and imputation purposes, as a vital
+element within the process of polar climate research.
+Sea ice forecasting
+The prediction of sea ice extent is an important task that in-
+forms safe shipping routes, hazard alerts, and climate pre-
+diction models (Wang et al. 2016). Predictions can even be
+used to issue warnings prior to events like massive haul-
+outs of walruses, providing the opportunity to prevent high
+mortality of the species (Andersson et al. 2021). Interan-
+nual variability makes sea ice forecasting a challenging task
+(Gagn ´e, Gillett, and Fyfe 2015; Andersson et al. 2021).
+Gagn ´e, Gillett, and Fyfe (2015) investigate the contrary re-
+sulting trends of simulated and actually observed sea ice data
+in the Antarctic by extending the historic records with recov-
+ered satellite based estimates from 35 to 50 years. The ad-
+ditional data further highlight the presence of high historic
+variability in the phenomenon, but emphasizes the view that
+existing climate simulations do not holistically describe the
+behaviour of sea ice extent. An application at the opposite
+end of the globe, the Beaufort Sea in the Arctic, uses convo-
+lutional neural networks (CNNs) to estimate high-resolution
+ice concentration maps directly from satellite synthetic aper-
+ture radar (SAR) data (Wang et al. 2016). SAR remote sens-
+ing is not impaired by cloud cover or the absence of day-
+light and is therefore a robust input. Although the regional
+scale of this application is constricted and sea ice concen-
+tration is not predicted for the future, the resulting perfor-
+mance, ranking close to the human expert benchmark, is
+a promising outcome. Since then, various researchers have
+applied deep learning models to predict sea ice concentra-
+tions, however for short, sub-seasonal lead times: Chi and
+Kim (2017) use deep learning and Kim et al. (2020) later
+use Convolutional Neural Networks (CNNs), a variant of
+deep learning, to predict Artic sea ice concentrations. Ali
+et al. (2022) propose an attention-based Long Short TermMemory (LSTM) ensemble method, combining the strength
+of attention-based methods to learn distant connections and
+the ability of LSTMs to remember previous states, analog
+to previous weather conditions. Ali et al.’s model outper-
+forms previous state-of-the-art models. However, these ap-
+plications only evaluate 1-month ahead predictions.
+In more recent work Andersson et al. (2021) present a
+machine learning model to predict monthly averaged sea
+ice probability classes across the entire Arctic region at
+lead times of 1 to 6 months. They use a range of differ-
+ent input data, including climate variables from the atmo-
+sphere and ocean. The model is constructed as an ensem-
+ble of U-Nets, a variant of CNNs. U-Nets were originally
+developed for biomedical image segmentation, a conceptu-
+ally similar Computer Vision task, mapping from gridded
+inputs (e.g. images) to gridded outputs. Andersson et al.’s
+IceNet model outperforms the state-of the art physics-based
+model at longer prediction lead times. The deep learning en-
+semble performs particularly well on predicting extreme sea
+ice conditions. Andersson et al.’s work is exemplary in in-
+tegrating domain knowledge and machine learning: It not
+only displays a high level of understanding for the domain,
+but it also extracts interpretable results from the model,
+that may in turn provide new insights to domain experts and
+their models. A variable importance analysis is used to un-
+derstand what inputs are contributing most to yield the pre-
+dictive results for different months and lead times. The find-
+ings are compared to expectations from sea ice forecasting
+experts, and are mostly found to match domain knowledge.
+Nonetheless some new discoveries were also made from this
+data-driven approach. One interesting result is that exten-
+sively pre-training the model on CMIP6 climate simulation
+data barely increased the predictive performance. This sup-
+ports the recognition that relatively small amounts of obser-
+vational data, rather than large amounts of simulated data,
+can be highly indicative of future phenomena when used
+within suitable modelling settings. Andersson et al. (2021)
+suggest extending their work by using inputs at higher tem-
+poral resolution, with the intention of improving predictive
+ability at short, 1-month, lead times, where the model is
+currently under-performing. Furthermore, the authors sug-
+gest incorporating ice thickness as a model input to further
+improve forecasts. While the classifier predicts a discrete
+probability distribution over the possible sea ice probability
+classes as an output, there are opportunities to expand on the
+methodological approach by incorporating the probabilistic
+framework.
+Climate feedbacks and teleconnections
+Teleconnections are persistent patterns of climate anomalies
+that span large geographical areas. Such patterns and their
+causal structures are hard to detect but they influence climate
+processes at the global scale. The work by Kretschmer et al.
+(2016) demonstrates an application of causal hypothesis
+testing to understand Arctic teleconnection patterns: Causal
+effect networks (CEN), a type of graphical model, are used
+on time series data to identify autumn Barents and Kara sea
+ice concentrations as an important driver for mid-latitude
+winter circulation, which can show as extreme winter condi-
+
+tions in North America and Euroasia. Artic teleconnections
+are currently not very well understood, and as identified by
+Rolnick et al. (2019) incorporating them into climate mod-
+els is likely to improve climate projections at global and re-
+gional resolutions. Work by van Nes et al. (2015) uses an-
+other type of technique, convergent cross-mapping (CCM), a
+non-linear state-space method , to investigate causal feed-
+back structures in the field of paleoclimatology (van Nes
+et al. 2015). They use more than 400,000 years of temper-
+ature data and greenhouse gas concentrations reconstructed
+from the V ostok Ice core from Antarctica as a proxy time
+series. Their results demonstrate that orbital forcing (e.g.
+insolation) have no significant causal association with ei-
+ther temperature or greenhouse gas concentrations. How-
+ever, they found a strong feedback effect of temperature vari-
+ability on greenhouse gases, indicating that warming in it-
+self may drive an increase in greenhouse gas concentrations.
+This constitutes an important finding on the level of cause
+and effect structures associated with climate change.
+Discussion, Opportunities and next steps
+In the following we discuss the applications reviewed in the
+previous section, and we examine their methods, and distill
+common challenges. Based on this we introduce and moti-
+vate opportunities for probabilistic Machine Learning meth-
+ods, in particular we introduce Bayesian Optimisation and
+Causal methods and what use cases for future work they pro-
+vide.
+Discussion
+Based on our review of recent works applying machine
+learning to the Polar Climate and Earth Science domain (see
+Table 1) we can observe that particular high-impact appli-
+cations, i.e. sea ice forecasting and sea level rise predic-
+tion, have received more attention than others. Deep learn-
+ing methods are often used in conjunction with satellite data,
+likely motivated by the success of deep learning for promi-
+nent Computer Vision problems, as well as recent achieve-
+ments in the cryosphere domain (Andersson et al. 2021;
+Ali et al. 2022). Currently, distributed sub-communities con-
+tribute to this new field and relevant work is published across
+research venues in Earth & Climate, Machine Learning &
+Statistics or interdisciplinary venues, where terminology,
+contribution emphasis, and reproducibiity standards vary.
+Challenges repeatedly discussed in the literature include:
+• combining data from various sources (data fusion);
+• dealing with varying spatial and temporal resolutions;
+• increasing computational efficiency;
+• interpretability of models and model outputs;
+• modelling natural variability of phenomena;
+• modelling systemic sources of uncertainty related to data
+and models.
+Andersson et al. (2021) showcase how variable importance
+analysis can be used for deep learning models to make sense
+of the mechanism behind the black-box-model to address in-
+terpretability. However, this work incorporates probabilistic
+representation only for predicted outputs. Zammit-Mangionet al. (2014) is one of the few to have used Bayesian statistics
+to model uncertainty throughout the hierarchical model; an
+endeavour calling for a high degree of domain expertise to
+inform prior distributions, parameterisation and model struc-
+ture. As uncertainty is inseparable from Climate research,
+there are major opportunities to use Probabilistic Machine
+Learning methods to solve the challanges faced.
+Opportunities for Probabilistic Machine Learning
+Probabilistic Machine Learning describes those methods
+that utilise a probabilistic framework to represent uncer-
+tainty. The probabilistic modelling framework is rooted in
+principled theoretical and highly practical approaches that
+are concerned with “representing and manipulating uncer-
+tainty” (Ghahramani 2015). Uncertainty arises from incor-
+rect or biased measurements, from decisions about model
+structure, from model parameters and from the stochastic
+nature of the world. Therefore, uncertainty should be propa-
+gated through the model and included in model predictions.
+A review paper (Ghahramani 2015) provides an excellent
+introduction to Bayesian inference, the core of Bayesian
+statistics, and an overview of recent advances, specifically,
+Bayesian Optimisation , probabilistic programming, prob-
+abilistic data compression, and automatic model discovery.
+Ghahramani highlights the importance of the probabilistic
+modelling framework for problems where uncertainty is a
+“key ingredient”. The paper also discusses a common com-
+putational challenge among these probabilistic methods -
+inference - and how approximate integration methods like
+Markov Chain Monte Carlo (MCMC) (refer to Andrieu et al.
+(2003); Brooks et al. (2011) for more detail) or Variational
+Inference (refer to Blei, Kucukelbir, and McAuliffe (2017))
+are related research fields addressing this challenge.
+Various methods featured in Ghahramani (2015), includ-
+ing the aforementioned Bayesian Optimisation and its most
+common underlying surrogate model, Gaussian Processes,
+originated from the spatial and spatio-temporal mod-
+elling literature. Gaussian Process regression, which is also
+known as Kriging in geostatistics, is a class of flexible
+non-parametric models that has been particularly success-
+ful in modelling spatial correlation structures (Marchant and
+Ramos 2012). Gaussian Process models are discussed in
+great detail in the textbook (Rasmussen and Williams 2006).
+In the area of spatio-temporal modelling, the seminal text-
+books by Cressie (2011, 1993) combine classical statisti-
+cal methods and modern computational algorithms and are
+therefore influential across theoretical and applied fields.
+Other methods which have been gaining scholarly popu-
+larity and are thus worth mentioning are Bayesian Neu-
+ral Networks (Chandra, Azizi, and Cripps 2017), which
+combine standard Neural Networks with Bayesian Infer-
+ence, and Causal Inference (refer to Pearl (2009)), which
+is concerned with ascertaining causal relationships using
+probabilistic tools. As reviewed in the previous section,
+Kretschmer et al. (2016) and van Nes et al. (2015) demon-
+strate the use of causal methods in climate studies. Next, we
+will discuss Bayesian Optimisation and Causal Inference, as
+well as opportunities for applying these to polar climate re-
+search, in more detail.
+
+Bayesian Optimisation. Bayesian Optimisation is a tool
+for global optimisation. It is particularly suitable when the
+objective function is unknown and complex, and when eval-
+uations of the objective function are noisy and costly to
+obtain (Marchant and Ramos 2012; Archetti and Cande-
+lieri 2019; Shahriari et al. 2016). Over iterations, each new
+query point, where the objectively function is then evalu-
+ated, will be chosen carefully and efficiently. While some
+applications focus on finding the global optimum, other ap-
+plication focus on the iterative determination of the next op-
+timal query point, known as active learning (Shahriari et al.
+2016). Well-known use cases for Bayesian Optimisation ex-
+ist in the design of exploration strategies for mining and ge-
+ology in environmental applications, where Bayesian Opti-
+misation can inform the design of sensing networks (Shahri-
+ari et al. 2016). Exploration drilling is a way of evaluating
+the unknown objective function, which describes the dis-
+tribution of sub-surface minerals across space. Exploration
+drilling is very costly. Hence, data-efficient Bayesian Opti-
+misation is well suited to inform decision making about the
+selection of promising drilling sites. In environmental mon-
+itoring Bayesian Optimisation is used to inform optimal se-
+quential decisions which result in efficient data acquisition
+of environmental variables of concern (Marchant, Ramos,
+and Sanner 2014). Bayesian Optimisation takes into account
+the expected value based on the global model (for the ‘opti-
+misation’ in Bayesian Optimisation) and also the degree of
+uncertainty the model has with regard to the expected value,
+based on the data, the underlying model assumptions and
+the prior. This is the trade-off between exploitation and ex-
+ploration. To take into account the added desiderata of min-
+imising sensor travel distance (Marchant and Ramos 2012)
+propose a new acquisition function, the Distance Based Up-
+per Confidence Bound. They demonstrate considerably re-
+duced travel distance in a real world and a simulated exper-
+iment without sacrificing accuracy. Use cases in polar re-
+search have strong parallels to this work, with limited sens-
+ing resources available, a vast space to explore, and high-
+uncertainty models. Because Bayesian Optimisation simul-
+taneously updates the probabilistic model of the unknown
+function and sequentially suggests sampling locations (ac-
+tive learning), the method has dual utility. Therefore, meth-
+ods building on top of these ideas, for example reflecting
+geographic or other asymmetric constraints in the acquisi-
+tion strategy, may be a possible extension of previous work
+with high practical relevance for Polar Climate research.
+Recent work has applied Bayesian Optimisation to ac-
+tively monitor urban air pollution in London using Hierar-
+chical Bayesian modelling as the surrogate model (Hellan,
+Lucas, and Goddard 2022). Further work is suggested to
+explore the use of other kernel families and kernel varia-
+tions that can capture correlations appearing at different time
+scales. Another application of Bayesian Optimisation to the
+environmental domain is the localisation of a contamination
+source (Pirot et al. 2019). This work provides a good exam-
+ple for integrating hydrology domain knowledge into the ob-
+jective function. Within the Machine Learning community,
+Bayesian Optimisation has attracted a lot of attention for
+its use in optimising hyperparameters of Machine Learningmodels (Snoek et al. 2014). Open-source software packages
+like Dragonfly (Kandasamy et al. 2020) enable a ready-to-
+use implementation of these ideas. Potentially this use case
+can be transferred to the optimisation of climate models, or
+to Machine Learning models of the Earth’s sub-systems.
+Expanding on existing research, future work could ap-
+ply Bayesian Optimisation to optimise sensor networks for
+climate monitoring in polar regions, or as an active learn-
+ing strategy to determine drilling locations for ice cores.
+An extension to the work of Marchant and Ramos (2012)
+could propose new acquisition functions that uses a non-
+stationary cost function which reflects the physical charac-
+teristics of the environment. Another challenging problem,
+shared across the reviewed literature (Gopalan, Zammit-
+Mangion, and McCormack 2021; Leong and Horgan 2020)
+is data fusion. Combining data from different remote sens-
+ing technologies as well as in situ measurements, demands
+a principled way of fusing varying uncertainty distributions,
+interpolating missing data or unifying scales. Whilst this is
+a sub-problem of applied research generally, the Bayesian
+framework may offer an elegant way to address this and
+therefore could benefit other applications of Data Science to
+climate-related domain problems in the Arctic and Antarc-
+tic.
+Causal Inference. Climate Modelling is predominately
+associated with prediction through the implementation of
+deterministic physical systems which are highly inter-
+pretable. With the rise in machine learning methods, a size-
+able component of the research community has focused
+on developing predictive black-box models that can be
+deployed as flexible and accurate regression (e.g. neural
+networks, computer vision, recommender systems). These
+methods, under the Rolnick et al. (2019) framework, are
+attributed to informing adaptation strategies in response to
+consequences of Climate Change. However, these methods
+present serious limitations from a scientific perspective since
+they: i) do not provide interpretability, thus limiting the ca-
+pacity for climate scientists to learn from model predictions;
+ii) show a lack of transparency into the underlying working
+of the models, which may lead to a lack of trust; and iii)
+capture correlations and not causation which may result in
+misleading and incorrect recommendations.
+In contrast, modelling the causal mechanisms of Cli-
+mate Change, thereby discerning anthropogenic and natural
+causes of warming, will provide insights that inform mit-
+igation strategies with stronger and interpretable evidence.
+Understanding the causes of phenomena we observe lies at
+the very heart of scientific discovery (Runge et al. 2019).
+Many domains, like medicine, use controlled experiments
+to establish causal links. However, in a large and complex
+field like Earth Science, where controlled experiments are
+impossible or unethical, Causal Inference methods based on
+observational data are a promising new research direction.
+Runge et al. provide an overview of Causal Inference frame-
+works for dealing with observational time series data and
+they suggest suitable applications in the Earth System Sci-
+ences. Computer simulation experiments, the prior standard
+for causal discovery in the Earth Sciences, are computation-
+
+ally expensive and constrained to assumptions made about
+the systems. Concurrent with the rise in Machine Learn-
+ing, data availability and increased computing power paved
+the way for these new causal methods, which rely only on
+observational data. Research in Bayesian Networks (Pearl
+2009) dates back a few decades but forms the foundation for
+many causal models. An important framework reviewed in
+Runge et al. (2019) is Structural Causal Models (SCMs)
+(refer to Peters, Janzing, and Sch ¨olkopf (2017) for more de-
+tail). These are closely related to Bayesian Networks. Both
+are graphical models, where the nodes of the graph repre-
+sent variables of interest and the links between nodes repre-
+sent causal relationships. SCMs are a particularly appealing
+framework, because various strong assumptions (e.g. about
+the noise structure) that were previously unavoidable, can
+be relaxed. SCMs can be viewed as a complement to black
+box ML models, to increase understanding of the mecha-
+nisms of the system (Runge et al. 2019). This understanding
+of causal relationships is not just a means to an end, but has
+also been recognised to increase robustness, particularly for
+out-of-distribution predictions (Runge et al. 2019).
+In the context of probabilistic machine learning and un-
+certainty quantification, the recent rise in fully probabilis-
+tic Bayesian network inference has the power of incorporat-
+ing uncertainty about causal structures by providing poste-
+rior distributions over graph structures (Kuipers and Moffa
+2017). Furthermore, if causal inference and causal effects
+are also treated in a fully probabilistic framework, they have
+the capacity to quantify uncertainty and guide sequential de-
+cision making. Causal inference can also be connected with
+Bayesian Optimisation (Aglietti et al. 2020), which can be
+generalised to active sampling and intervention strategies
+that acquire data in order to find the most valuable actions.
+Novel developments in Causal Inference frameworks in-
+cluding Bayesian Networks andStructural Causal Mod-
+elsenable us to gain understanding of causal structures of
+underlying systems from observational data. These offer
+great opportunities for future work, for instance, to build
+more robust climate models, to further understand causal
+feedbacks in climate change as demonstrated by van Nes
+et al. (2015), or to distinguish anthropogenic from natural
+drivers of Climate Change.
+Conclusion
+Opportunities for applying Machine Learning to solve prob-
+lems from the Climate Sciences and the Polar Climate Sci-
+ences more specifically, are widely recognised and have
+the potential to be highly impactful (Rolnick et al. 2019).
+However, because this interdisciplinary research area is still
+novel, and remote sensing data has only become more ac-
+cessible and more meaningful with increased sensing cov-
+erage and accompanying computing power in recent years,
+there are more research opportunities than existing work. A
+large body of work exists on the use of deep learning for
+remote sensing application (Ma et al. 2019) as well as the
+Earth Sciences (Reichstein et al. 2019). Aligned with this,
+many reviewed applications of machine learning to polar cli-
+mate research use deep learning in combination with satel-
+lite data. Some of these applications outperform state-of-the art physics-based models (Andersson et al. 2021), sug-
+gesting further promising advances in this direction of re-
+search in the future. Common challenges across reviewed
+literature include the need for data fusion, assimilating
+multi-resolution data, increasing computational efficiency,
+enhancing interpretability, and modelling uncertainty. Ad-
+dressing these challenges is another opportunity for future
+work and will benefit research down-stream.
+Although probabilistic modelling is inevitable for making
+sensible and informed decisions, methods applied to prob-
+lems in this fields often lack a framework for uncertainty
+quantification. To address this need, the class of probabilis-
+tic Machine Learning offers a toolbox of methods which
+are well-suited to reflect real-life uncertainty. We particu-
+larly highlight Bayesian Optimisation and Causal Inference
+methods which are well suited to problems from the Polar
+Climate and Earth Science domain. Bayesian Optimisation
+may be used to inform drilling site selection of ice cores, se-
+quential selection of monitoring locations for autonomous
+sensors, or to optimise stationary sensor networks across
+the polar regions. Other non-spatial applications include the
+global optimisation of hyperparameters for machine learn-
+ing and traditional climate models. Improved experimental
+design may help in reducing the computational footprint of
+this computationally intensive field of research. Advances
+in Causal Inference techniques provide another great op-
+portunity for future work: Quantifying causal drivers of cli-
+mate change or building more robust prediction models, by
+resembling the underlying causal structures of the system,
+could strengthen the uncertainty-aware, scientific founda-
+tion for global decision making in stewarding human im-
+pact on climate, thereby supporting climate change mitiga-
+tion and adaption efforts.
+References
+Abraham, R.; and Ueda, Y . 2000. The Chaos Avant-garde:
+Memories of the Early Days of Chaos Theory . World Scien-
+tific. ISBN 978-981-238-647-2.
+Aglietti, V .; Lu, X.; Paleyes, A.; and Gonz ´alez, J. 2020.
+Causal Bayesian Optimization. In Proceedings of the Twenty
+Third International Conference on Artificial Intelligence
+and Statistics , 3155–3164. PMLR. ISSN: 2640-3498.
+Ali, S.; Huang, Y .; Huang, X.; and Wang, J. 2022. Sea
+Ice Forecasting using Attention-based Ensemble LSTM.
+arXiv:2108.00853 . ArXiv:2108.00853.
+Andersson, T. R.; Hosking, J. S.; P ´erez-Ortiz, M.; Paige,
+B.; Elliott, A.; Russell, C.; Law, S.; Jones, D. C.; Wilkin-
+son, J.; Phillips, T.; Byrne, J.; Tietsche, S.; Sarojini, B. B.;
+Blanchard-Wrigglesworth, E.; Aksenov, Y .; Downie, R.; and
+Shuckburgh, E. 2021. Seasonal Arctic sea ice forecasting
+with probabilistic deep learning. Nature Communications ,
+12(1): 5124.
+Andrieu, C.; de Freitas, N.; Doucet, A.; and Jordan, M. I.
+2003. An Introduction to MCMC for Machine Learning.
+Machine Learning , 50(1): 5–43.
+Archetti, F.; and Candelieri, A. 2019. Bayesian Optimization
+and Data Science . SpringerBriefs in Optimization. Cham:
+
+Springer International Publishing. ISBN 978-3-030-24493-
+4 978-3-030-24494-1.
+Balaji, V .; Maisonnave, E.; Zadeh, N.; Lawrence, B. N.;
+Biercamp, J.; Fladrich, U.; Aloisio, G.; Benson, R.; Caubel,
+A.; Durachta, J.; Foujols, M.-A.; Lister, G.; Mocavero, S.;
+Underwood, S.; and Wright, G. 2017. CPMIP: measure-
+ments of real computational performance of Earth system
+models in CMIP6. Geoscientific Model Development , 10(1):
+19–34. Publisher: Copernicus GmbH.
+Blei, D. M.; Kucukelbir, A.; and McAuliffe, J. D. 2017. Vari-
+ational Inference: A Review for Statisticians. Journal of the
+American Statistical Association , 112(518): 859–877.
+Brooks, S.; Gelman, A.; Jones, G.; and Meng, X.-L. 2011.
+Handbook of Markov Chain Monte Carlo . London, United
+Kingdom: CRC Press LLC. ISBN 978-1-4200-7942-5.
+Chandra, R.; Azizi, L.; and Cripps, S. 2017. Bayesian Neu-
+ral Learning via Langevin Dynamics for Chaotic Time Se-
+ries Prediction. In Liu, D.; Xie, S.; Li, Y .; Zhao, D.; and
+El-Alfy, E.-S. M., eds., Neural Information Processing , Lec-
+ture Notes in Computer Science, 564–573. Cham: Springer
+International Publishing. ISBN 978-3-319-70139-4.
+Chi, J.; and Kim, H.-c. 2017. Prediction of Arctic Sea Ice
+Concentration Using a Fully Data Driven Deep Neural Net-
+work. Remote Sensing , 9(12): 1305.
+CIRES. 2022. Sea Level Research Group.
+Https://sealevel.colorado.edu/. Accessed: 2022-07-20.
+Climate Change AI. 2022. Climate Change AI - About.
+Https://www.climatechange.ai/. Accessed: 2022-07-20.
+Copernicus. 2021. Latest projections of future climate now
+available. Https://climate.copernicus.eu/latest-projections-
+future-climate-now-available. Accessed: 2022-07-20.
+Cressie, N. 1993. Statistics for Spatial Data . New York,
+United States: John Wiley & Sons, Incorporated. ISBN 978-
+1-119-11517-5.
+Cressie, N. A. C. 2011. Statistics for spatio-temporal data .
+Wiley series in probability and statistics. Hoboken, N.J: Wi-
+ley. ISBN 978-0-471-69274-4.
+CSIRO. 2016. Sea-level Rise: CSIRO & ACE-
+CRC. Https://www.cmar.csiro.au/sealevel/sl about us.html.
+Accessed: 2022-07-20.
+Dirscherl, M.; Dietz, A. J.; Kneisel, C.; and Kuenzer, C.
+2020. Automated Mapping of Antarctic Supraglacial Lakes
+Using a Machine Learning Approach. Remote Sensing ,
+12(7): 1203.
+Elsevier. 2022. Polar Science Aims and scope.
+Https://www.sciencedirect.com/journal/polar-
+science/about/aims-and-scope. Accessed: 2022-07-20.
+Eyring, V .; Bony, S.; Meehl, G. A.; Senior, C. A.; Stevens,
+B.; Stouffer, R. J.; and Taylor, K. E. 2016. Overview
+of the Coupled Model Intercomparison Project Phase 6
+(CMIP6) experimental design and organization. Geosci-
+entific Model Development , 9(5): 1937–1958. Publisher:
+Copernicus GmbH.
+Fretwell, P.; Pritchard, H. D.; Vaughan, D. G.; Bamber,
+J. L.; Barrand, N. E.; Bell, R.; Bianchi, C.; Bingham, R. G.;
+Blankenship, D. D.; Casassa, G.; Catania, G.; Callens, D.;Conway, H.; Cook, A. J.; Corr, H. F. J.; Damaske, D.;
+Damm, V .; Ferraccioli, F.; Forsberg, R.; Fujita, S.; Gim, Y .;
+Gogineni, P.; Griggs, J. A.; Hindmarsh, R. C. A.; Holmlund,
+P.; Holt, J. W.; Jacobel, R. W.; Jenkins, A.; Jokat, W.; Jor-
+dan, T.; King, E. C.; Kohler, J.; Krabill, W.; Riger-Kusk, M.;
+Langley, K. A.; Leitchenkov, G.; Leuschen, C.; Luyendyk,
+B. P.; Matsuoka, K.; Mouginot, J.; Nitsche, F. O.; Nogi, Y .;
+Nost, O. A.; Popov, S. V .; Rignot, E.; Rippin, D. M.; Rivera,
+A.; Roberts, J.; Ross, N.; Siegert, M. J.; Smith, A. M.; Stein-
+hage, D.; Studinger, M.; Sun, B.; Tinto, B. K.; Welch, B. C.;
+Wilson, D.; Young, D. A.; Xiangbin, C.; and Zirizzotti, A.
+2013. Bedmap2: improved ice bed, surface and thickness
+datasets for Antarctica. The Cryosphere , 7(1): 375–393.
+Publisher: Copernicus GmbH.
+Gagn ´e, M.-E.; Gillett, N. P.; and Fyfe, J. C. 2015. Observed
+and simulated changes in Antarctic sea ice extent over the
+past 50 years. Geophysical Research Letters , 42(1): 90–95.
+Ghahramani, Z. 2015. Probabilistic machine learning and
+artificial intelligence. Nature , 521(7553): 452–459. Num-
+ber: 7553 Publisher: Nature Publishing Group.
+Gopalan, G.; Zammit-Mangion, A.; and McCormack, F.
+2021. A Review of Bayesian Modelling in Glaciology.
+arXiv:2112.13663 [stat] .
+Hellan, S. P.; Lucas, C. G.; and Goddard, N. H. 2022.
+Bayesian Optimisation for Active Monitoring of Air Pollu-
+tion. arXiv:2202.07595 [physics] . ArXiv: 2202.07595.
+IMBIE. 2022. IMBIE. Http://imbie.org/. Accessed: 2022-
+07-20.
+IPCC. 2019. Summary for Policymakers. In P ¨ortner, H.-
+O.; D.C. Roberts; V . Masson-Delmotte; P. Zhai; M. Tignor;
+E. Poloczanska; K. Mintenbeck; A. Alegr ´ıa; M. Nicolai; A.
+Okem; J. Petzold; B. Rama; and N.M. Weyer, eds., IPCC
+Special Report on the Ocean and Cryosphere in a Chang-
+ing Climate , pp. 3–35. Cambridge, UK and New York, NY ,
+USA: Cambridge University Press.
+Kandasamy, K.; Vysyaraju, K. R.; Neiswanger, W.; Paria,
+B.; Collins, C. R.; Schneider, J.; P ´oczos, B.; and Xing, E. P.
+2020. Tuning hyperparameters without grad students: scal-
+able and robust Bayesian optimisation with dragonfly. The
+Journal of Machine Learning Research , 21(1): 81:3098–
+81:3124.
+Kim, Y . J.; Kim, H.-C.; Han, D.; Lee, S.; and Im, J. 2020.
+Prediction of monthly Arctic sea ice concentrations using
+satellite and reanalysis data based on convolutional neural
+networks. The Cryosphere , 14(3): 1083–1104. Publisher:
+Copernicus GmbH.
+Kopp, R. E.; DeConto, R. M.; Bader, D. A.; Hay, C. C.;
+Horton, R. M.; Kulp, S.; Oppenheimer, M.; Pollard, D.; and
+Strauss, B. H. 2017. Evolving Understanding of Antarctic
+Ice-Sheet Physics and Ambiguity in Probabilistic Sea-Level
+Projections. Earth’s Future , 5(12): 1217–1233.
+Kretschmer, M.; Coumou, D.; Donges, J. F.; and Runge, J.
+2016. Using Causal Effect Networks to Analyze Different
+Arctic Drivers of Midlatitude Winter Circulation. Journal of
+Climate , 29(11): 4069–4081. Publisher: American Meteo-
+rological Society Section: Journal of Climate.
+
+Kuipers, J.; and Moffa, G. 2017. Partition MCMC
+for Inference on Acyclic Digraphs. Journal of
+the American Statistical Association , 112(517):
+282–299. Publisher: Taylor & Francis eprint:
+https://doi.org/10.1080/01621459.2015.1133426.
+Leong, W.; and Horgan, H. 2020. DeepBedMap: A deep
+neural network for resolving the bed topography of Antarc-
+tica. Cryosphere , 14(11): 3687–3705.
+Lythe, M. B.; and Vaughan, D. G. 2001. BEDMAP: A
+new ice thickness and subglacial topographic model of
+Antarctica. Journal of Geophysical Research: Solid Earth ,
+106(B6): 11335–11351.
+Ma, L.; Liu, Y .; Zhang, X.; Ye, Y .; Yin, G.; and Johnson,
+B. A. 2019. Deep learning in remote sensing applications:
+A meta-analysis and review. ISPRS Journal of Photogram-
+metry and Remote Sensing , 152: 166–177.
+Marchant, R.; and Ramos, F. 2012. Bayesian optimisa-
+tion for Intelligent Environmental Monitoring. In 2012
+IEEE/RSJ International Conference on Intelligent Robots
+and Systems , 2242–2249. ISSN: 2153-0866.
+Marchant, R.; Ramos, F.; and Sanner, S. 2014. Sequen-
+tial Bayesian optimisation for spatial-temporal monitoring.
+InProceedings of the Thirtieth Conference on Uncertainty
+in Artificial Intelligence , UAI’14, 553–562. Arlington, Vir-
+ginia, USA: AUAI Press. ISBN 978-0-9749039-1-0.
+Monteleoni, C.; Schmidt, G. A.; and McQuade, S. 2013. Cli-
+mate Informatics: Accelerating Discovering in Climate Sci-
+ence with Machine Learning. Computing in Science & En-
+gineering , 15(5): 32–40.
+NASA. 2021. Landsat 1 |Landsat Science.
+Https://landsat.gsfc.nasa.gov/satellites/landsat-1/. Ac-
+cessed: 2022-07-20.
+Naughtin, C.; Hajkowicz, S.; Schleiger, E.; Bratanova, A.;
+Cameron, A.; Zamin, T.; and Dutta, A. 2022. Our Future
+World: Global megatrends impacting the way we live over
+coming decades. Technical report, CSIRO, Brisbane, Aus-
+tralia. Publisher: CSIRO.
+NOAA. 2022. Explore Sea Level Rise
+Tools, Services, and Educational Material.
+Https://oceanservice.noaa.gov/hazards/sealevelrise/. Ac-
+cessed: 2022-07-20.
+Notz, D.; and Community, S. 2020. Arctic Sea Ice
+in CMIP6. Geophysical Research Letters , 47(10):
+e2019GL086749.
+Pearl, J. 2009. Causality: models, reasoning, and inference .
+Cambridge ;: Cambridge University Press, 2nd ed. edition.
+ISBN 978-0-521-89560-6.
+Peters, J.; Janzing, D.; and Sch ¨olkopf, B. 2017. Elements
+of Causal Inference: Foundations and Learning Algorithms .
+Adaptive Computation and Machine Learning series. Cam-
+bridge, MA, USA: MIT Press. ISBN 978-0-262-03731-0.
+Pirot, G.; Krityakierne, T.; Ginsbourger, D.; and Renard, P.
+2019. Contaminant source localization via Bayesian global
+optimization. Hydrology and Earth System Sciences , 23(1):
+351–369. Publisher: Copernicus GmbH.Rasmussen, C. E.; and Williams, C. K. I. 2006. Gaussian
+processes for machine learning . Adaptive computation and
+machine learning. Cambridge, Mass: MIT Press. ISBN 978-
+0-262-18253-9.
+Reichstein, M.; Camps-Valls, G.; Stevens, B.; Jung, M.;
+Denzler, J.; Carvalhais, N.; and Prabhat. 2019. Deep learn-
+ing and process understanding for data-driven Earth system
+science. Nature , 566(7743): 195–204.
+Rolnick, D.; Donti, P. L.; Kaack, L. H.; Kochanski, K.; La-
+coste, A.; Sankaran, K.; Ross, A. S.; Milojevic-Dupont, N.;
+Jaques, N.; Waldman-Brown, A.; Luccioni, A.; Maharaj, T.;
+Sherwin, E. D.; Mukkavilli, S. K.; Kording, K. P.; Gomes,
+C.; Ng, A. Y .; Hassabis, D.; Platt, J. C.; Creutzig, F.; Chayes,
+J.; and Bengio, Y . 2019. Tackling Climate Change with
+Machine Learning. arXiv:1906.05433 [cs, stat] . ArXiv:
+1906.05433.
+Runge, J.; Bathiany, S.; Bollt, E.; Camps-Valls, G.; Coumou,
+D.; Deyle, E.; Glymour, C.; Kretschmer, M.; Mahecha,
+M. D.; Mu ˜noz-Mar ´ı, J.; van Nes, E. H.; Peters, J.; Quax, R.;
+Reichstein, M.; Scheffer, M.; Sch ¨olkopf, B.; Spirtes, P.; Sug-
+ihara, G.; Sun, J.; Zhang, K.; and Zscheischler, J. 2019. In-
+ferring causation from time series in Earth system sciences.
+Nature Communications , 10(1): 2553.
+Shahriari, B.; Swersky, K.; Wang, Z.; Adams, R. P.; and
+de Freitas, N. 2016. Taking the Human Out of the Loop:
+A Review of Bayesian Optimization. Proceedings of the
+IEEE , 104(1): 148–175. Conference Name: Proceedings of
+the IEEE.
+Shi, J.-R.; Talley, L. D.; Xie, S.-P.; Peng, Q.; and Liu, W.
+2021. Ocean warming and accelerating Southern Ocean
+zonal flow. Nature Climate Change , 11(12): 1090–1097.
+Shirmard, H.; Farahbakhsh, E.; M ¨uller, R. D.; and Chandra,
+R. 2022. A review of machine learning in processing re-
+mote sensing data for mineral exploration. Remote Sensing
+of Environment , 268.
+Snoek, J.; Swersky, K.; Zemel, R.; and Adams, R. 2014. In-
+put Warping for Bayesian Optimization of Non-Stationary
+Functions. In Proceedings of the 31st International Con-
+ference on Machine Learning , 1674–1682. PMLR. ISSN:
+1938-7228.
+Springer Nature. 2022. Climate sciences.
+Https://www.nature.com/subjects/climate-sciences. Ac-
+cessed: 2022-07-20.
+van Nes, E. H.; Scheffer, M.; Brovkin, V .; Lenton, T. M.; Ye,
+H.; Deyle, E.; and Sugihara, G. 2015. Causal feedbacks in
+climate change. Nature Climate Change , 5(5): 445–448.
+Wang, L.; Scott, K. A.; Xu, L.; and Clausi, D. A. 2016. Sea
+Ice Concentration Estimation During Melt From Dual-Pol
+SAR Scenes Using Deep Convolutional Neural Networks: A
+Case Study. IEEE Transactions on Geoscience and Remote
+Sensing , 54(8): 4524–4533.
+Zammit-Mangion, A.; Rougier, J.; Bamber, J.; and Sch ¨on,
+N. 2014. Resolving the Antarctic contribution to sea-level
+rise: a hierarchical modelling framework. Environmetrics ,
+25(4): 245–264.
+
+Zammit-Mangion, A.; Rougier, J.; Sch ¨on, N.; Lindgren, F.;
+and Bamber, J. 2015. Multivariate spatio-temporal mod-
+elling for assessing Antarctica’s present-day contribution to
+sea-level rise. Environmetrics , 26(3): 159–177.
+Acknowledgments
+This research was supported by the Australian Government
+through the Australian Research Council’s Industrial Trans-
+formation Training Centre in Data Analytics for Resources
+and Environments (DARE) (project IC190100031). Further-
+more, this research was supported by an Australian Govern-
+ment Research Training Program (RTP) Scholarship.
+

aaaifss2022_17.txt

@@ -0,0 +1,315 @@
+Rethinking Machine Learning for Climate Science: A Dataset Perspective
+Aditya Grover1,2
+1Department of Computer Science
+2Institute of the Environment and Sustainability
+University of California, Los Angeles
+Abstract
+The growing availability of data sources is a predominant
+factor enabling the widespread success of machine learning
+(ML) systems across a wide range of applications. Typically,
+training data in such systems constitutes a source of ground-
+truth , such as measurements about a physical object (e.g.,
+natural images) or a human artifact (e.g., natural language).
+In this position paper, we take a critical look at the validity
+of this assumption for datasets for climate science. We argue
+that many such climate datasets are uniquely biased due to the
+pervasive use of external simulation models (e.g., general cir-
+culation models) and proxy variables (e.g., satellite measure-
+ments) for imputing and extrapolating in-situ observational
+data. We discuss opportunities for mitigating the bias in the
+training and deployment of ML systems using such datasets.
+Finally, we share views on improving the reliability and ac-
+countability of ML systems for climate science applications.
+1 Introduction
+Large datasets are fueling major advances in the scal-
+ing of machine learning (ML) systems for a variety of real-
+world usecases of relevance to science and society, ranging
+from creative art and text generation (Ramesh et al. 2021;
+Brown et al. 2020) to protein folding (Jumper et al. 2021)
+and drug discovery (Vamathevan et al. 2019). This has led
+to a growing optimism for the broad field of climate change
+as well (Rolnick et al. 2022). With advancements in sensory,
+storage, and network technology, we now have large datasets
+available for many domains of interest to climate change,
+such as weather forecasting (Rasp et al. 2020), agriculture
+and forestry (Zheng et al. 2019), and chemical and materi-
+als discovery (Kirklin et al. 2015; Chanussot et al. 2021),
+among others.
+As the first step of any ML pipeline, the choice of a train-
+ing dataset is critical to the downstream performance of ML
+systems. Both the quantity and quality of a dataset play an
+important role, as demonstrated by numerous prior studies
+(e.g., (Gebru et al. 2021)) that correlate the size, noise, and
+bias within training datasets with broad and holistic indica-
+tors of downstream performance, such as accuracy and fair-
+Copyright © 2022, Association for the Advancement of Artificial
+Intelligence (www.aaai.org). All rights reserved.ness. Given the growing enthusiasm in using ML for cli-
+mate change, it begs the question: are datasets for climate
+domains aligned with ML pipelines in use today?
+In this position paper, we argue that climate science do-
+mains can present unique challenges for ML systems given
+how datasets are collected and generated. In particular, we
+note that climate datasets used in practice are routinely based
+onreanalysis orgridding that combine disparate real and
+simulated/proxy measurement sources. While such a proce-
+dure ensures that datasets have excellent coverage, it leads to
+a bias that can propagate within standard ML pipelines. This
+calls for a rethink of both training and deployment of data-
+centric ML pipelines for climate science, as well as commu-
+nity guidelines for dataset and model release.
+The rest of the paper is structured as follows: in Section 2,
+we briefly review current data collection practices in climate
+science and the role of ML in improving climate projections.
+In Section 3, we present opportunities for aligning machine
+learning with data practices in climate science, as well as
+community guidelines for improving the transparency and
+accountability of ML models. Finally, we conclude in Sec-
+tion 4 with a summary and discussion on broader impacts,
+including implications of this research on domains focusing
+on climate change mitigation and adaptation, as well as other
+disciplines within ML.
+2 What Makes Climate Data Unique?
+Climate modeling is fundamental to understanding the inter-
+actions between the atmospheric, oceanic, and land surface
+process, including anthropogenic interventions. Such mod-
+els can be used for short-term weather forecasts or long-term
+projections of the Earth’s climate under different interven-
+tions. Beyond scientific pursuits, the outputs of these models
+inform regional and international policy aimed at near- and
+long-term climate mitigation and adaptation.
+Typically, climate models couple our physical under-
+standing with on-ground observations. However, such mod-
+els can be insufficient for certain downstream usecases due
+to limited accuracy and/or spatiotemporal resolution. For ex-
+ample, nowcasting requires very short-horizon weather pre-
+dictions (up to 2 hours ahead) that is greater than the time
+it takes to spin up numerical weather prediction (NWP) sys-
+tems (Ravuri et al. 2021). Similarly, many general circula-
+tion models (GCM) and earth system models (ESM) that are
+
+used for projecting future climate operate at a 2 degree reso-
+lution ( 200km), which is much lower than typically needed
+(<0.1 degrees) for effective mitigation planning at a regional
+level (Fowler, Blenkinsop, and Tebaldi 2007).
+In such scenarios, data-driven solutions involving ma-
+chine learning can play a big role in overcoming the limi-
+tations of current climate models. However, the quality of a
+ML system depends significantly on the availability of high
+quality datasets. This presents two key challenges. First, his-
+torical in-situ observational records for climate variables are
+irregularly sampled due to uneven access to sensory tech-
+nology, leading to a geographical bias. Second, for climate
+change in particular, we require projections of future climate
+under different interventions (e.g., different fossil fuel us-
+age) — many of these scenarios have never been observed
+in the past, but are necessary for governments and interna-
+tional organizations to analyze and formulate policies.
+Together, the above challenges necessitate the use of al-
+ternate data sources, such as reanalysis datasets andgrid-
+ded datasets . Reanalysis datasets combine historical obser-
+vations with the outputs of climate models, whereas grid-
+ded datasets rely on statistical tools for imputing miss-
+ing values or proxy measurements made via satellites. In
+both cases, the goal is to generate high volume and high
+coverage datasets for training ML systems. Several such
+datasets are in use today, such as CHIRPS (Funk et al. 2015),
+a gridded dataset for high-resolution rainfall combining
+satellite measurements with in-situ observations, and ERA-
+5 (Mu ˜noz-Sabater et al. 2021), a reanalysis dataset main-
+tained by the European Centre for Medium-Range Weather
+Forecasts. These datasets are updated daily and contain his-
+torical observations spanning many decades, providing ex-
+cellent spatiotemporal coverage at the expense of their re-
+spective model bias. As a concrete example, consider data
+for soil moisture available from the ERA reanalysis dataset.
+Soil moisture is an important climate variable for project-
+ing the agriculture viability of any land area. For validation
+on real measurements, ERA5 uses in-situ soil measurement
+data from 14 sites — 4 in North America, 6 in Europe, 1 in
+Australia, and 2 in Africa, reflecting a highly biased distri-
+bution with respect to global demographics and completely
+omitting some continents.
+3 Roadmap for Climate ML Pipelines
+In the previous section, we motivated the use of reanalysis
+and gridded datasets for training ML models, and the inher-
+ent bias they encode. How should we train ML systems on
+such climate datasets? The status quo, as adopted in several
+papers (e.g., Oses et al. (2020); Ba ˜no-Medina, Manzanas,
+and Guti ´errez (2020)), is to treat the reanalysis dataset as
+ground-truth. However, this ignores the context in which the
+dataset was generated and is likely to propagate or even po-
+tentially amplify the bias in the dataset. While there is no
+simple solution, we believe that ML pipelines that explicitly
+account for this additional context can be far more effec-
+tive for downstream applications. In this regard, we outline
+our position on exciting directions for improving the training
+and deployment of ML pipelines for climate science.3.1 Training
+Model selection. While training benefits immensely from
+the use of high coverage (but biased) datasets, we can con-
+sider alternate strategies for model selection (e.g., via the
+use of validation datasets). In areas for which we have in-
+situ observations, we can monitor the model’s performance
+directly on such data for the held-out years, sidestepping any
+bias due to the use of gridded or reanalysis tools. Also, note
+that since model selection is less data-hungry than training
+the model itself, this strategy can also be potentially applied
+for underserved regions with few in-situ measurements.
+Unsupervised learning and domain adaptation. In the
+last few years, there have been several advances in large
+scale unsupervised representation learning, including both
+contrastive and generative approaches (Murphy 2022).
+While in-situ measurements of climate variables are hard to
+obtain for arbitrary targets, we can obtain high-quality fea-
+ture descriptors for unsupervised pretraining.
+Alternatively, a closely related problem is that of unsu-
+pervised domain adaptation, where we need to transfer ML
+models trained on one domain to a related domain (with zero
+or few labels). Various techniques have been developed to
+enable such a transfer, such as the use of domain randomiza-
+tion (Tobin et al. 2017) for control tasks. In the climate con-
+text, we can consider the gridded/reanalysis datasets as the
+source domain and consider transferring ML models trained
+on such datasets to points in the target domain of interest.
+3.2 Deployment
+Uncertainty quantification. Well-calibrated uncertainty
+estimates can play a key role in reliably communicating the
+predictions of ML systems trained on gridded and reanal-
+ysis datasets and downstream users relying on theses pre-
+dictions. In principle, one could use any gridded or reanaly-
+sis dataset for training a ML model. However, as one might
+expect, different datasets differ in their imputation strate-
+gies and hence, the predictions of ML models trained on
+these datasets would also differ. Consequently, we can treat
+these models as an ensemble (Lakshminarayanan, Pritzel,
+and Blundell 2017) and use the distribution of predictions
+for each of the ML models as a measure of uncertainty due
+to the imputation strategy.
+Datasheets and model cards. While the need for docu-
+menting datasets and models is well-recognized in both the
+ML and climate communities, the standards and terminolo-
+gies vary significantly. As we see more real-world deploy-
+ments, it is important to expand the scope of existing pro-
+tocols, such as datasheets (Gebru et al. 2021) and model
+cards (Mitchell et al. 2019) in the ML community, to better
+document key details relating to the gridded and reanalyzed
+datasets, such as the details on the auxiliary climate models
+and data sources used for dataset creation, the distribution of
+in-situ measurement sites, and any known limitations of the
+imputation strategy. We believe including such details can
+significantly improve the transparency and interpretability
+of ML systems, as well as aid in reproducibility — a grow-
+ing area of concern for ML in scientific applications (Kapoor
+and Narayanan 2022).
+
+4 Broader Impacts
+This position paper calls for a careful reflection on the use of
+datasets for ML applications in climate science. We argued
+that while current reanalysis and gridded datasets might
+seem to have global coverage at high spatiotemporal band-
+widths, these datasets are in fact reflective of the geographic
+and socioeconomic disparities in access to sensory technol-
+ogy (e.g., satellites, weather balloons). Quantifying and mit-
+igating this bias without compromising on overall accuracy
+is an open challenge for the ML community. Our work high-
+lights a select group of directions in this regard grounded in
+metrics concerning accuracy, reliability, and reproducibility.
+While the use of gridded and reanalysis datasets is com-
+mon practice in the climate science community, we also ex-
+pect similar challenges in other fields relevant to climate
+change, and ML more broadly. For example, efforts to use
+ML for computational chemistry are fundamentally bottl-
+necked by the domain gap in computational simulation soft-
+wares and real experimental data. Even more so, with the ad-
+vent and rapid proliferation of deep generative models, we
+are likely to find future ML systems trained on mixtures of
+real and synthetic data, and thus leading to a natural cross-
+pollination of tools and techniques.
+References
+Ba˜no-Medina, J.; Manzanas, R.; and Guti ´errez, J. M. 2020.
+Configuration and intercomparison of deep learning neural
+models for statistical downscaling. Geoscientific Model De-
+velopment , 13(4): 2109–2124.
+Brown, T.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J. D.;
+Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell,
+A.; et al. 2020. Language models are few-shot learners.
+NeurIPS , 33: 1877–1901.
+Chanussot, L.; Das, A.; Goyal, S.; Lavril, T.; Shuaibi, M.;
+Riviere, M.; Tran, K.; Heras-Domingo, J.; Ho, C.; Hu, W.;
+et al. 2021. Open catalyst 2020 (OC20) dataset and commu-
+nity challenges. ACS Catalysis , 11(10): 6059–6072.
+Fowler, H. J.; Blenkinsop, S.; and Tebaldi, C. 2007. Link-
+ing climate change modelling to impacts studies: recent ad-
+vances in downscaling techniques for hydrological mod-
+elling. International Journal of Climatology: A Journal of
+the Royal Meteorological Society , 27(12): 1547–1578.
+Funk, C.; Peterson, P.; Landsfeld, M.; Pedreros, D.; Verdin,
+J.; Shukla, S.; Husak, G.; Rowland, J.; Harrison, L.; Hoell,
+A.; et al. 2015. The climate hazards infrared precipitation
+with stations—a new environmental record for monitoring
+extremes. Scientific data , 2(1): 1–21.
+Gebru, T.; Morgenstern, J.; Vecchione, B.; Vaughan, J. W.;
+Wallach, H.; Iii, H. D.; and Crawford, K. 2021. Datasheets
+for datasets. Communications of the ACM , 64(12): 86–92.
+Jumper, J.; Evans, R.; Pritzel, A.; Green, T.; Figurnov, M.;
+Ronneberger, O.; Tunyasuvunakool, K.; Bates, R.; ˇZ´ıdek,
+A.; Potapenko, A.; et al. 2021. Highly accurate protein struc-
+ture prediction with AlphaFold. Nature , 596(7873): 583–
+589.
+Kapoor, S.; and Narayanan, A. 2022. Leakage and the Re-
+producibility Crisis in ML-based Science.Kirklin, S.; Saal, J. E.; Meredig, B.; Thompson, A.; Doak,
+J. W.; Aykol, M.; R ¨uhl, S.; and Wolverton, C. 2015. The
+Open Quantum Materials Database (OQMD): assessing the
+accuracy of DFT formation energies. npj Computational
+Materials , 1(1): 1–15.
+Lakshminarayanan, B.; Pritzel, A.; and Blundell, C. 2017.
+Simple and scalable predictive uncertainty estimation using
+deep ensembles. NeurIPS , 30.
+Mitchell, M.; Wu, S.; Zaldivar, A.; Barnes, P.; Vasserman,
+L.; Hutchinson, B.; Spitzer, E.; Raji, I. D.; and Gebru, T.
+2019. Model cards for model reporting. In Proceedings
+of the conference on fairness, accountability, and trans-
+parency , 220–229.
+Mu˜noz-Sabater, J.; Dutra, E.; Agust ´ı-Panareda, A.; Al-
+bergel, C.; Arduini, G.; et al. 2021. ERA5-Land: A state-of-
+the-art global reanalysis dataset for land applications. Earth
+System Science Data , 13(9): 4349–4383.
+Murphy, K. P. 2022. Probabilistic machine learning: an in-
+troduction . MIT press.
+Oses, N.; Azpiroz, I.; Marchi, S.; Guidotti, D.; Quartulli, M.;
+and G. Olaizola, I. 2020. Analysis of copernicus’ era5 cli-
+mate reanalysis data as a replacement for weather station
+temperature measurements in machine learning models for
+olive phenology phase prediction. Sensors , 20(21): 6381.
+Ramesh, A.; Pavlov, M.; Goh, G.; Gray, S.; V oss, C.; Rad-
+ford, A.; Chen, M.; and Sutskever, I. 2021. Zero-shot text-to-
+image generation. In International Conference on Machine
+Learning , 8821–8831. PMLR.
+Rasp, S.; Dueben, P. D.; Scher, S.; Weyn, J. A.; Mouatadid,
+S.; and Thuerey, N. 2020. WeatherBench: a benchmark data
+set for data-driven weather forecasting. Journal of Advances
+in Modeling Earth Systems , 12(11): e2020MS002203.
+Ravuri, S.; Lenc, K.; Willson, M.; Kangin, D.; Lam, R.;
+Mirowski, P.; Fitzsimons, M.; Athanassiadou, M.; Kashem,
+S.; Madge, S.; et al. 2021. Skilful precipitation nowcasting
+using deep generative models of radar. Nature , 597(7878):
+672–677.
+Rolnick, D.; Donti, P. L.; Kaack, L. H.; Kochanski, K.; La-
+coste, A.; Sankaran, K.; Ross, A. S.; Milojevic-Dupont, N.;
+Jaques, N.; Waldman-Brown, A.; et al. 2022. Tackling cli-
+mate change with machine learning. ACM Computing Sur-
+veys (CSUR) , 55(2): 1–96.
+Tobin, J.; Fong, R.; Ray, A.; Schneider, J.; Zaremba, W.;
+and Abbeel, P. 2017. Domain randomization for transfer-
+ring deep neural networks from simulation to the real world.
+In2017 IEEE/RSJ international conference on intelligent
+robots and systems (IROS) , 23–30. IEEE.
+Vamathevan, J.; Clark, D.; Czodrowski, P.; Dunham, I.; Fer-
+ran, E.; Lee, G.; Li, B.; Madabhushi, A.; Shah, P.; Spitzer,
+M.; et al. 2019. Applications of machine learning in drug
+discovery and development. Nature reviews Drug discov-
+ery, 18(6): 463–477.
+Zheng, Y .-Y .; Kong, J.-L.; Jin, X.-B.; Wang, X.-Y .; Su, T.-
+L.; and Zuo, M. 2019. CropDeep: the crop vision dataset for
+deep-learning-based classification and detection in precision
+agriculture. Sensors , 19(5): 1058.
+

aaaifss2022_18.txt

@@ -0,0 +1,783 @@
+Graph Representation Learning for Energy Demand Data:
+Application to Joint Energy System Planning under Emissions Constraints
+Aron Brenner*1, Rahman Khorramfar*2, Dharik Mallapragada3, Saurabh Amin4
+1,4Civil and Environmental Engineering (CEE) and Laboratory for Information &Decision Systems (LIDS)
+2MIT Energy Initiative (MITEI) and Laboratory for Information &Decision Systems (LIDS)
+3MIT Energy Initiative (MITEI)
+{abrenner, khorram, dharik, amins }@mit.edu
+Abstract
+A rapid transformation of current electric power and natural
+gas (NG) infrastructure is imperative to meet the mid-century
+goal of CO 2emissions reduction requires. This necessitates
+a long-term planning of the joint power-NG system under
+representative demand and supply patterns, operational con-
+straints, and policy considerations. Our work is motivated by
+the computational and practical challenges associated with
+solving the generation and transmission expansion problem
+(GTEP) for joint planning of power-NG systems. Specifi-
+cally, we focus on efficiently extracting a set of representa-
+tive days from power and NG data in respective networks and
+using this set to reduce the computational burden required
+to solve the GTEP. We propose a Graph Autoencoder for
+Multiple time resolution Energy Systems (GAMES) to cap-
+ture the spatio-temporal demand patterns in interdependent
+networks and account for differences in the temporal resolu-
+tion of available data. The resulting embeddings are used in a
+clustering algorithm to select representative days. We evalu-
+ate the effectiveness of our approach in solving a GTEP for-
+mulation calibrated for the joint power-NG system in New
+England. This formulation accounts for the physical interde-
+pendencies between power and NG systems, including the
+joint emissions constraint. Our results show that the set of
+representative days obtained from GAMES not only allows
+us to tractably solve the GTEP formulation, but also achieves
+a lower cost of implementing the joint planning decisions.
+Introduction
+One of the most significant societal challenges that we cur-
+rently face is to transition to a reliable, low-carbon, and
+sustainable energy system as soon as possible, and to meet
+the mid-century goal of limiting global warming below 2◦C
+(UN-FCCC 2015; Gielen et al. 2019). This requires a signif-
+icant use of renewable energy resources and well-planned
+integration of various energy vectors, including emerging
+clean energy sources such as hydrogen and other renew-
+able energy sources. Our work is motivated by the enor-
+mous potential of machine learning (ML) models in pro-
+moting sustainable energy systems. In particular, we focus
+on ML modeling for extracting a set of representative days
+from heterogeneous demand data associated with real-world
+*The first two authors contributed equally to this work.
+Copyright © 2022, Association for the Advancement of Artificial
+Intelligence (www.aaai.org). All rights reserved.electric power and natural gas (NG) systems, and using this
+set for joint power-NG network planning under emissions
+constraints. In doing so, we leverage ML-extracted repre-
+sentative days to tractably solve an optimization problem
+that determines a capacity and network expansion plan for
+regional-scale energy systems such as that of New England.
+Broadly speaking, our work addresses several practi-
+cal and computational challenges associated with capac-
+ity expansion models (CEMs) for decarbonization of in-
+terdependent power-NG infrastructures. Classical examples
+of such models include the generation expansion problem
+(GEP) and generation and transmission expansion problem
+(GTEP), both of which are well-studied in the context of
+power systems (Li et al. 2022a; He et al. 2018). Our opti-
+mization model is a GTEP that determines the optimal lo-
+cation and timing of generation units, transmission lines,
+and pipelines to meet future energy demands under a range
+of operational and policy constraints such as joint emission
+constraints. In our work, we extend the model to include
+two main interdependencies between power and NG sys-
+tems. The first interdependency captures the increasing role
+of gas-fired power plants in the generation mix of electricity
+production ( EIA; He et al. 2018). The second interdepen-
+dency reflects the joint emission of CO 2in both systems.
+The key computational challenge in solving the GTEP
+arises from the fact that it links long-term investment de-
+cisions (e.g. capacity and network expansion) to short-term
+operational ones (e.g. unit commitment, power production,
+and energy storage). The former decisions have a planning
+horizon of 10-30 years with yearly granularity, while the
+latter usually require hourly or sub-hourly resolution. Un-
+der reasonable assumptions, we can express the GTEP as a
+large-scale mixed-integer linear program (MILP), but cur-
+rent literature has limited success in tractably solving these
+problems to an adequate level of spatial and temporal resolu-
+tion. In our case, the computational difficulty in solving the
+GTEP increases further because we model both power and
+NG networks. Thus, taking into account (projected) demand
+information on a day-to-day basis becomes prohibitively
+expensive from a computational viewpoint. In the classi-
+cal GTEP problems for power systems, the computational
+challenge is addressed by aggregating power system nodes
+(buses) within a geographical neighborhood (power zone) to
+a single node (Li et al. 2022a) and by solving the GTEP for a
+
+set of representative days (Hoffmann et al. 2020). Crucially,
+the set of representative days needs to capture demand and
+supply patterns. To the best of our knowledge, the notion of
+representative days has not been clearly defined and devel-
+oped in the context of joint power-NG planning problem –
+this is where we leverage our graph representation learning
+approach.
+Our work also addresses the practical issues arising from
+coarse data availability from the NG network. Firstly, we do
+not have access to the detailed connectivity and transmis-
+sion information in the NG network while this information
+is readily available for the power network. Secondly, power
+systems typically collect demand and generation data at a
+fine temporal resolution (hourly or less), but this data is usu-
+ally not publicly accessible for NG systems. These issues
+thus require us to (a) formulate network constraints based on
+loosely specified information on power and NG node con-
+nectivity and (b) develop an approach to leverage demand
+and supply data from the power system with demand data of
+NG system despite their different temporal resolutions.
+We address the aforementioned challenges by develop-
+ing a graph representation learning approach that captures
+the physical interdependencies between power and NG net-
+works, and also handles the different granularity of data
+at each network. We consider demand data for both sys-
+tems, and consider capacity factor (CF) data for solar and
+wind plants to reflect the supply pattern in the renewable-
+dominated future grid. We utilize graph convolutions to cap-
+ture the network interactions both within and across power
+and NG networks, and adopt an autoencoder architecture
+with tuneable reconstruction losses for the respective de-
+mand and CF data. We demonstrate that the resulting Graph
+Autoencoder for Multiple time resolution Energy Sys-
+tems (GAMES) model is ideally suited to handle embed-
+ding the spatio-temporal patterns in power and NG demand
+as well as wind and solar CF data into a lower-dimensional
+representation, which can be readily clustered to extract the
+set of representative days. Furthermore, our approach to
+computing the set of representative days can also enable an
+accurate estimation of the trade-off between costs (both in-
+vestment and operational) and joint emissions from power
+and NG systems.1
+Previous studies for selecting representative days propose
+variants of k-means (Mallapragada et al. 2018; Li et al.
+2022b; Teichgraeber and Brandt 2019; Barbar and Mallapra-
+gada 2022), k-medoids (Scott et al. 2019; Teichgraeber and
+Brandt 2019), and hierarchical clustering (Liu, Sioshansi,
+and Conejo 2017; Teichgraeber and Brandt 2019). The dis-
+tance matrices used in clustering algorithms for most previ-
+ous works are constructed based on a set of time series inputs
+such as load data and variable renewable energies (VRE)
+capacity factors (Li et al. 2022a; Hoffmann et al. 2020).
+Notably, these approaches neither account for demand data
+with multiple time resolutions nor account for network in-
+1We believe this capability can have a significant societal im-
+pact by lowering the barriers to investment in renewable energy re-
+sources and alleviating reliability concerns in a low-carbon energy
+system.terdependencies. Hence, they cannot be readily extended to
+address the task of extracting representative days for joint
+power-NG systems – an aspect that is crucial for realism and
+tractability in joint planning optimization models for decar-
+bonizing these systems. We believe that our GAMES model
+addresses these challenges and provides a promising path
+to better extract representative days in interdependent power
+and NG systems.
+Graph Convolutional Autoencoder Approach
+In this section, we describe the Graph Autoencoder for Mul-
+tiple time resolution Energy Systems (GAMES) model, a
+simple graph autoencoder with linear graph convolutions.
+We argue that this architecture efficiently captures spatio-
+temporal demand patterns in power and NG systems.
+Autoencoders
+To begin with, we note that direct use of clustering algo-
+rithms to identify representative days for any large-scale en-
+ergy system is prone to the “curse of dimensionality” due
+to the high dimensionality of time series data. In such set-
+tings, it is desirable to first extract low-dimensional and de-
+noised representations of the data before clustering (Par-
+sons, Haque, and Liu 2004). To identify a set of represen-
+tative days, we choose to utilize a state-of-the-art autoen-
+coder architecture for learning low-dimensional embeddings
+for power-NG systems (that have different time resolutions)
+prior to clustering.
+Given a high-dimensional input such as a time series of
+graph signals, X∈Rp, an autencoder can be trained to
+jointly learn an encoder, g:Rp→Rk, and a decoder,
+f:Rk→Rpthat minimize the reconstruction loss func-
+tion∥X−ˆX∥2
+2, where ˆX=f(g(X))is the reconstructed
+signal. Here, k≪pdenotes the dimension of the learned
+latent space.
+Variable Interpretation Granularity Nodes
+XE Electricity Hourly 188
+XW Wind Hourly 188
+XS Solar Hourly 188
+XG Natural Gas Daily 18
+Table 1: Notation for input variables.
+We denote by XE∈Rd×nE×tEthe data tensor of elec-
+tricity demands over all days d, nodes nE, and times tE.
+Similarly, we denote the natural gas data tensor by XG∈
+Rd×nG×tG, the wind capacity factor tensor by XW∈
+Rd×nW×tW, and the solar capacity factor data tensor by
+XS∈Rd×nS×tS(see Table 1). Because the GTEP considers
+different associated costs for investment and operational de-
+cisions related to power, NG, wind, and solar, we introduce
+hyperparameters αG, αW, αSin the autoencoder objective
+function to tune the trade-off between the multiple recon-
+struction losses. This parameter reflects the contribution of
+each system towards the total cost. For example, if the NG
+system cost is twice the power system cost, then higher val-
+ues of αGensure that the reconstruction cost is penalized
+
+05101520Hour°2.0°1.5°1.0°0.50.0Electricity Demand (std. dev.)Node 8Node 205101520Hour°2.0°1.5°1.0°0.50.0Electricity Demand (std. dev.)Node 72Node 92
+05101520Hour°2.0°1.5°1.0°0.50.0Electricity Demand (std. dev.)Node 142Node 13005101520Hour°2.0°1.5°1.0°0.50.0Electricity Demand (std. dev.)Node 170Node 155Figure 1: Adjacent nodes in the power network demonstrate similar variations in demand over the course of the day. These
+spatial dependencies are modeled explicitly by graph convolutional layers in the GAMES architecture.
+more when deviating from the data of the NG system. This
+gives us the following loss function:
+dX
+i=11
+dnEtE∥X(i)
+E−ˆX(i)
+E∥2
+F+αG
+dnGtG∥X(i)
+G−ˆX(i)
+G∥2
+F
++αW
+dnWtW∥X(i)
+W−ˆX(i)
+W∥2
+F+αS
+dnStS∥X(i)
+S−ˆX(i)
+S∥2
+F
+,
+where ∥ · ∥Fdenotes the Frobenius norm.
+In our case study, we set αG= 2, αS= 0.5, αW= 0.5.
+However, we note that it is possible to choose the hyperpa-
+rameters by evaluating the downstream GTEP objective for
+different values. Specifically, this can be performed using a
+grid search in which the quality of a combination of hyper-
+parameters {αG, αW, αS}is measured by GTEP objective
+costs given by solving the optimization model rather than
+the GAMES validation loss directly.
+Graph Representation Learning
+Next, we provide a brief introduction to modeling with graph
+convolutional networks (GCNs).Preliminaries We encode the network topology with the
+binary adjacency2matrix A, which we construct such that
+Aij=0 (i, j)/∈ E
+1 (i, j)∈ E.
+We also construct the diagonal degree matrix Dsuch that
+Dii=P
+jAij.
+Graph Convolutions Our graph autoencoder approach
+follows (Kipf and Welling 2017) in utilizing Chebyshev con-
+volutional filters , which approximate spectral convolutions
+to learn node embeddings as weighted local averages of em-
+beddings of adjacent nodes. This is ideal for learning low-
+dimensional embeddings of energy networks as neighbor-
+hoods of nodes typically exhibit similar energy demands
+patterns and can thus be represented jointly. Chebyshev fil-
+ters operate on the “renormalized” graph Laplacian ˜L=
+˜D−1
+2˜A˜D−1
+2, where ˜D=I+Dand˜A=I+A, and
+perform a form of Laplacian smoothing (Li, Han, and Wu
+2018; Taubin 1995). We initialize H(0)=Xand apply con-
+volutional filters to learn subsequent node embeddings as
+2Ideally, one should construct an affinity matrix Awith a Gaus-
+sian kernel such that Aij= exp
+−dist(i,j)2
+σ2
+for all edges (i, j),
+where dist(i, j)denotes the distance of edge (i, j)andσdenotes
+the standard deviation of distances in the network (Shuman et al.
+2012). Since we do not have access to edge distance data in our
+case study, we proceed with the binary adjacency matrix.
+
+follows:
+H(l+1)=σ(˜LH(l)Θ(l)),
+where Θ(l)is a trainable weight matrix and H(l)is a matrix
+of node embeddings in layer l.σ(·)is typically a nonlinear
+activation function, such as ReLU ortanh .
+In each layer, GCNs aggregate features from the imme-
+diate neighborhood of each node. Deep GCNs stack multi-
+ple layers with nonlinear activations to learn node embed-
+dings as nonlinear functions of both local and global node
+features. In contrast, (Salha, Hennequin, and Vazirgiannis
+2019) propose a simpler graph autoencoder model, which
+they demonstrate to have competitive performances with
+multilayer GCNs on standard benchmark datasets despite
+being limited to linear first-order interactions. Shallow neu-
+ral architectures are also better suited for settings where data
+is scarce. This is particularly significant in modeling energy
+systems whose data may only be available for a few his-
+torical years. Indeed, we find this simpler GCN approach to
+perform well for our case study. We now introduce GAMES,
+an augmented version of the linear GCN autoencoder for en-
+ergy systems with multiple time resolutions.
+GAMES
+Our proposed GAMES architecture is designed as follows
+and illustrated in Fig. 2.
+Encoder Consider the power, wind CF, solar CF, and NG
+time series corresponding to day i,X(i)
+E,X(i)
+W,X(i)
+S,X(i)
+G.
+We begin by constructing the data matrix X(i)as
+X(i)= 
+X(i)
+EX(i)
+WX(i)
+W 0
+0 0 0 X(i)
+G!
+.
+Note that X(i)∈Rn×t, where n:=nE+nGandt:=
+tE+tW+tW+tG. This is because capacity factor data
+exists for all nodes in the power network and utilizes the
+same network topology. X(i)is then passed through a sin-
+gle convolutional layer to produce the low-dimensional em-
+bedding Z(i)∈Rn×k. The hyperparameter kdefines the
+bottleneck of the autoencoder architecture (i.e. the dimen-
+sion of each node embedding) and consequently the tradeoff
+between compression and reconstruction loss. In our case
+study, we find k= 3 to show a sufficient performance for
+our application of identifying representative days.
+Decoder Z(i)is passed through a convolutional layer to
+produce the embedding H(i)∈Rn×t. This reconstructed
+matrix is then split along the second dimension into two
+blocks: H(i)
+E,W,S∈R(nE+nW+nS)×tandH(i)
+G∈RnG×t.
+Each block is then passed to a separate series of fully con-
+nected layers with tanh activations that map the node em-
+beddings in H(i)
+E,W,SandH(i)
+Grespectively to the reconstruc-
+tions ˆX(i)
+E,W,SandˆX(i)
+G. Finally, the tensor ˆX(i)
+E,W,Sis split
+into the reconstructions ˆX(i)
+E,ˆX(i)
+W,ˆX(i)
+S.Clustering
+After the model is trained, the power
+and NG time series from each day i, i.e.
+(X(1)
+E, X(1)
+W, X(1)
+S, X(1)
+G), . . . , (X(N)
+E, X(N)
+W, X(N)
+S, X(N)
+G),
+is passed through the encoder to generate the embedding
+matrices Z(1), . . . , Z(N). Then, k-medoids clustering is
+applied to select a set of Kcluster medians, denoted by
+S ⊂ { 1, . . . , N }, and assign each day ito a corresponding
+cluster j∈ S. We denote the set of days assigned to the
+cluster defined by day jasCj. Given the number of clusters
+K, the k-medoids algorithm aims to minimize the objective
+function
+minX
+j∈SX
+i∈Cj∥Z(i)−Z(j)∥2
+F (1)
+(Hastie, Tibshirani, and Friedman 2001). Note that every day
+in the dataset must be assigned to exactly one cluster. Se-
+mantically, (1) can be understood as aiming to ensure that
+the set of representative days Sproportionately partitions
+the full set of days in the dataset by minimizing squared Eu-
+clidean distances in the latent space as constructed by the
+autoencoder.
+Capacity Expansion Model
+The result of the clustering algorithm is used to solve the
+CEM for joint power and NG planning, which is formulated
+as a GTEP. The problem determines the minimum invest-
+ment cost and operational decisions for the year 2050 un-
+der various investment, operational, and policy constraints.
+The investment decisions include establishing new power
+plants, transmission lines, and pipelines as well as decom-
+missioning existing plants. The operational constraints in-
+clude minimum production, ramping, energy balance, trans-
+mission, and storage. We consider emission limits and min-
+imum share of VREs as policy constraints. Importantly, in
+our formulation, the emissions constraint limits CO 2emis-
+sions incurred by the consumption of NG in both networks.
+We introduce our model with simplified notation in this
+section and provide a detailed formulation in the supple-
+mentary material (SI 2022). Let ze= (xe,ye,p)represent
+the set of variables for the power system. The integer vari-
+ablexeis the variable establishing plants, decommissioning
+plants, and establishing new transmission lines. The contin-
+uous variable pcaptures the power generation in NG-fired
+plants while yeis a continuous variable that captures all the
+remaining variables including power generation from non
+NG-fired plants and power flow between nodes, storage, and
+load shedding variables. We use zg= (xg,yg,f)to de-
+note the set of variables associated with the NG system. The
+mixed-integer variable xgis the set of all investment, stor-
+age, and load shedding decisions. The continuous variable
+ygrepresents the intra-network flow, i.e. the flow between
+NG nodes or the flow between NG nodes and NG storage
+facilities. The flow between NG and electricity systems is
+denoted by f. We formulate the joint power-NG system as
+
+[XE∥XW∥XS]
+XGX ConvEncoder
+Z ConvDecoder
+HHE,W,S
+HGFC tanh FC [ˆXE∥ˆXW∥ˆXS]
+FC tanh FC ˆXG
+Extract Node EmbeddingsFigure 2: The GAMES Architecture. The electric power, wind CF, solar CF, and NG time series are combined into the block
+matrix Xwithnrows and tE+tW+tS+tGchannels. A single linear graph convolutional layer constructs matrix Zby
+embedding each row of Xintokdimensions. Another graph convolutional layer scales each row of Zback to tE+tW+tS+tG
+dimensions, which are then separated and fed through fully connected layers to reconstruct the two time series. After the model
+is trained, the embeddings are extracted by feeding the daily time series inputs through the encoder, at which point clustering is
+applied.
+follows:
+min (ce
+1xe+ce
+2ye+ce
+3p) + (cg
+1xg+cg
+2yg+cg
+3f)
+(2a)
+s.t.Aexe+Beye+Dep≤be
+1 (2b)
+Heye≥be
+2 (2c)
+Agxg+Bgyg+Dgf≤bg
+1 (2d)
+f=E1p (2e)
+G2yg+E2p≤η (2f)
+xe∈Z+,ye,xg∈Z+×R+,p,yg,f∈R+(2g)
+The objective function (2a) minimizes the investment and
+operational costs for the power system (first term) and NG
+system (second term). The constraint (2b) represents all in-
+vestment, commitment, and operational constraints for the
+power system including the production limit, ramping, stor-
+age, and energy balance constraints. The constraint (2c) en-
+forces policy considerations such as the minimum require-
+ment for renewable portfolio standard (RPS). The NG con-
+straints are reflected in constraint (2d), which includes tech-
+nological and operational constraints such as the supply
+limit at each node, flow between NG nodes, and storage.
+The coupling constraint (2e) ensures that NG-fired plants
+operate based on the gas flow they receive from the NG
+network. The second coupling constraint (2f) is the decar-
+bonization constraint that limits emissions resulting from
+NG consumption to serve both electricity (via NG power
+plants) and non-power related NG loads to η. The coeffi-
+cient matrices E1,G2, andE2represent the heat rate, emis-
+sion factors for NG usage, and emission factor for NG-fired
+plants, respectively. Indeed, emissions from coal-fired plants
+is a major driver for decarbonization efforts and NG remains
+as primary fuel for which emissions need to be regulated.
+Therefore, given the declining role of coal in the US energy
+system, the constraint (2f) reflects a futuristic setting where
+such plants are already decommissioned.Input Data
+Using publicly available data, we consider the New England
+region and construct its corresponding power and NG net-
+work. We then calibrate the resulting networks using his-
+torical data. The power network consists of 188 nodes with
+338 existing and candidate transmission lines. The NG net-
+work consists of 18 NG nodes and 7 storage nodes. We as-
+sume that each NG node is connected to two other storage
+nodes. We also assume that each power node is connected
+to three of its closest NG nodes. The Supplementary Infor-
+mation provides the details of the input data for the joint
+power-NG planning model (SI 2022).
+Computational Experiments
+GAMES Performance
+We train GAMES on a dataset of 292 days using the Adam
+optimizer with a learning rate of 0.001. We use the full batch
+of 292 data points for each update step and perform early
+stopping to end training when the validation loss no longer
+decreases. We report the validation reconstruction loss on
+a set of 73 days for various node embedding dimensions k
+in Table 2. We observe slightly diminishing returns for the
+Embed. Dim. k= 1 k= 2 k= 3 k= 4
+MSE Loss 0.727 0.398 0.244 0.160
+Table 2: The reconstruction loss shows diminishing returns
+fork >3node embedding dimensions.
+validation reconstruction loss for k >3. Consequently, we
+proceed with our representative day selection using embed-
+dings generated by the model corresponding to k= 3.
+Representative Days Comparison
+Setup We use the k-medoids clustering algorithm to ob-
+tain different sets of representative days. We apply the clus-
+tering algorithm to both raw data and the embeddings ob-
+
+2 6 10 14 18 22 26 30 34 38 421.52.01e10
+Total Cost
+Raw Data GAMES
+2 6 10 14 18 22 26 30 34 38 421.01.51e10
+Power System Cost
+2 6 10 14 18 22 26 30 34 38 424.55.01e9
+NG System Cost
+2 6 10 14 18 22 26 30 34 38 421.01.21e10
+Investment and FOM for Geneneration and Storage (Pow. Sys)
+2 6 10 14 18 22 26 30 34 38 420241e9
+Power System Load Shedding Cost
+2 6 10 14 18 22 26 30 34 38 42
+Number of Representative Days051e6
+Emission from Power System(a) GAMES vs. raw data clustering comparison under an 80%
+carbon reduction goal.
+2 6 10 14 18 22 26 30 34 38 421.752.002.251e10
+Total Cost
+Raw Data GAMES
+2 6 10 14 18 22 26 30 34 38 421.01.51e10
+Power System Cost
+2 6 10 14 18 22 26 30 34 38 42561e9
+NG System Cost
+2 6 10 14 18 22 26 30 34 38 421.01.21.41e10
+Investment and FOM for Geneneration and Storage (Pow. Sys)
+2 6 10 14 18 22 26 30 34 38 420241e9
+Power System Load Shedding Cost
+2 6 10 14 18 22 26 30 34 38 42
+Number of Representative Days051e6
+Emission from Power System(b) GAMES vs. raw data clustering comparison under a 95% car-
+bon reduction goal.
+Figure 3: Various costs and power emission for different number of representative days under different decarbonization goals.
+tained from the GAMES model to compare the results of
+the proposed model. Accordingly, two different sets are ob-
+tained for each number of representative days. The optimiza-
+tion model over the full power network is prohibitively chal-
+lenging even for a very small number of days. Therefore, we
+aggregate all buses in each state of the New England region
+to obtain a 6-node power network. This aggregation allows
+us to run the formulation for up to 42 representative days.
+We obtain a feasible solution in two steps for each set of
+representative days: (1) The optimization model is aggre-
+gated to the set of representative days for tractability and
+then solved. (2) Next, we consider the full planning horizon
+(the entire year of 2050) and set the integer decision vari-
+ables (i.e. investment decisions) to the values determined in
+the first step. We note that the investment decision variables
+in our formulation are (a) the only integer-valued decision
+variables and (b) independent of planning periods. There-
+fore, fixing them reduces the remaining operational problem
+to a linear program (LP), which can be solved considerably
+faster. The resulting solution from the second step is a fea-
+sible solution to the full-year problem, with which we can
+analyze resulting costs and decisions.In our computational experiments, we consider two de-
+carbonization goals of 80 %and 95 %where the former is the
+projected target for New England states (Weiss and Hagerty
+2019), and the latter aims reflects a radical decarbonization
+goal. Figures 3a and 3b show the results under 80 %and 95 %
+emission reduction goals respectively. Both figures evalu-
+ate the following quantities for the clusters obtained from
+GAMES and raw data: i) “Total Cost” which is the objec-
+tive function of model 2; ii) “Power System Cost” which is
+the first term in the objective function (2a); iii) “NG Sys-
+tem Cost” which is the second term in the objective func-
+tion (2a); iv) “Investment and FOM for Generation and Stor-
+age (Pow. Sysm)” (investment-FOM) which is part of the
+power system cost and captures the capital investment and
+fixed operating and maintenance (FOM) costs of installing
+new power plants and storage systems; v) “Power System
+Load Shedding Cost” which is part of the power system cost
+and reflects the cost of unsatisfied electricity demand; and
+v) “Emission from Power System” which is the tonnage of
+emission as a result of operating NG-fired power plants in
+the power system. We use “GAMES” to denote the feasi-
+ble solution for the set of days obtained by GAMES. We do
+
+Table 3: Average percentage change when using GAMES approach for for various costs and power emissions.
+Reduction Goal Total Power NG Inv-FOM (Power) Shedding cost (Power) Emission from Power Sys
+80% -5.14 -7.03 0.24 -4.64 -24.13 -9.87
+95% -7.27 -10.50 1.5 -8.51 -27.80 -3.31
+not report the wall-clock times, but all instances are solved
+under 5 hours. As expected, run-times vary significantly de-
+pending on the number of representative days utilized; in-
+stances with 2 representative days typically run in fewer than
+350 seconds, whereas 30-day instances may need to run for
+2800 seconds. All instances are implemented in Python us-
+ing Gurobi 9.5 and are run on the MIT Supercloud system
+with an Intel Xeon Platinum 8260 processor containing up
+to 96 cores and 192 GB of RAM (Reuther et al. 2018).
+Results Table 3 presents the percentage change in various
+quantities yielded by the GAMES representative days solu-
+tion as compared to the solution using representative days
+selected from clustering the raw data. The cost comparisons
+are also plotted in Figures 3a and 3b. We observe, on aver-
+age, a 5.14 %and 7.27 %improvement (decrease) in the total
+cost when using GAMES under 80 %and 95 %decarboniza-
+tion goals, respectively. This improvement may be attributed
+to GAMES’ ability to model dependencies between power
+and NG system data. Under more stringent decarbonization
+targets, the share of VRE increases and the role of dispatch-
+able power plants, such as NG-fired plants, diminishes. As
+a result, modeling the influence of capacity factors and their
+interactions with power and gas demands becomes more es-
+sential. This phenomenon may underlie our observation for
+the 22-day instance in which, while both approaches provide
+similar results under the 80 %decarbonization goal, GAMES
+significantly outperforms the raw data clustering as mea-
+sured by total cost for the higher decarbonization goal. As
+shown in Figures 3a and 3b, the total cost from GAMES out-
+performs or matches the performance of the raw data clus-
+tering in all instances (except the 30-day instance under an
+80%reduction goal). Interestingly, this disparity in perfor-
+mance is most drastic when 15 or fewer representative days
+are utilized under both decarbonization goals. This is worth
+noting as the optimization model instantiated on the full net-
+work topology (i.e. without aggregating nodes by state) is
+only tractable over a small set of representative days (i.e. af-
+ter applying a very coarse temporal aggregation). It is espe-
+cially important when the a model-year model only affords
+to consider a handful of representative days for each year.
+The power system cost largely drives variation in the to-
+tal cost under both decarbonization goals – the total cost is
+lower for all solutions with a lower power system cost. Note
+that the difference in performance is more pronounced in the
+power system cost compared to the total cost as indicated
+by the 7.03 %and 10.50 %power system cost improvement
+for GAMES under the 80 %and 95 %decarbonization goals.
+In Figure 3a, this trend aligns with load shedding costs ex-
+cept for the 14-day instance. However, as the 24.13 %de-
+crease shows, the GAMES approach results in significantly
+lower load shedding on average. The 27.80 %improvementin the load shedding cost for GAMES under the 95 %goal
+is plotted in detail in Figure 3b; GAMES outperforms the
+raw data clustering for all instances. Moreover, the GAMES
+approaches converges after 14 days with load shedding cost
+significantly lower than those instances utilizing fewer rep-
+resentative days.
+In both figures the trends of investment-FOM cost and
+power system cost are the same, indicating that the power
+system cost is largely driven by investment-FOM cost, and
+to a lesser extent, by load shedding cost. This is expected as
+future energy systems will rely significantly on VREs such
+as solar and wind power, which only incur investment and
+FOM costs. Another interesting observation pertains to the
+quantity of emissions in the power system caused by oper-
+ating NG-fired plants. Emissions for the power system are
+on average 9.87 %and 3.31 %lower for GAMES under the
+two decarbonization goals. This indicates a greater share
+of VREs in the GAMES approach, and correspondingly, a
+higher share of gas-fired plants in the raw data clustering
+approach. This is an interesting observation that may have
+significant implications for energy policy-making. In partic-
+ular, it suggests that the results from the raw data cluster-
+ing approach may be misleading as they do not sufficiently
+convey the radical changes required to transform the sys-
+tem from the current gas-dominant generation portfolio to a
+renewable-dominant power grid.
+NG system cost is another essential component of the total
+costs. Although NG costs are similar for GAMES and raw
+data clustering for each instance, the NG cost increases with
+the number of representative days. A possible explanation
+might be that neither GAMES nor raw data clustering aim to
+capture extreme days with separate clusters. Therefore, days
+with loads similar to extreme days are more likely to be se-
+lected as a cluster’s medoid as the number of representative
+days increases, which inevitably raises the NG system cost.
+This consideration is also consistent with the observed load
+shedding cost for the power system, which is significantly
+higher for instances with fewer than 15 representative days,
+indicating that both approaches fail to account for extreme
+days in cluster medoids.
+Conclusion
+In this work, we propose GAMES, a graph convolutional
+autoencoder for modeling energy demand in interdependent
+electric power and natural gas systems with heterogeneous
+nodes and different time resolutions. GAMES is able to ex-
+ploit spatio-temporal demand patterns to learn efficient em-
+beddings of interdependent power and NG networks. We ap-
+ply the k-medoids clustering algorithm to these embeddings
+to identify a set of representative days with which we are
+able to tractably solve an energy system infrastructure plan-
+
+ning problem calibrated for the joint power-NG system in
+New England. Our computational results show that the pro-
+posed framework outperforms clustering methods applied to
+the raw data and is effective in selecting a small number
+of representative days to provide high-quality feasible so-
+lutions for the optimization problem.
+The current work can be extended in multiple directions.
+The immediate extension of the GCN architecture is to ex-
+plore alternative approaches to graph representation learning
+such as Laplacian sharpening (Park et al. 2019). The extrac-
+tion and inclusion of extreme days, or low-frequency days
+with unusually low or high demand is another potential next
+step which could prevent high load shedding costs and better
+represents the NG system’s load patterns.
+References
+Barbar, M.; and Mallapragada, D. S. 2022. Represen-
+tative period selection for power system planning using
+autoencoder-based dimensionality reduction. arXiv preprint
+arXiv:2204.13608 .
+(EIA), E. I. A. 2022. EIA Website. Website. Accessed:
+2022-2-18.
+Gielen, D.; Gorini, R.; Wagner, N.; Leme, R.; Gutierrez, L.;
+Prakash, G.; Asmelash, E.; Janeiro, L.; Gallina, G.; Vale,
+G.; et al. 2019. Global energy transformation: a roadmap to
+2050.
+Hastie, T.; Tibshirani, R.; and Friedman, J. 2001. The Ele-
+ments of Statistical Learning . Springer Series in Statistics.
+New York, NY , USA: Springer New York Inc.
+He, C.; Zhang, X.; Liu, T.; Wu, L.; and Shahidehpour, M.
+2018. Coordination of interdependent electricity grid and
+natural gas network—a review. Current Sustainable/Renew-
+able Energy Reports , 5(1): 23–36.
+Hoffmann, M.; Kotzur, L.; Stolten, D.; and Robinius, M.
+2020. A review on time series aggregation methods for en-
+ergy system models. Energies , 13(3): 641.
+Kipf, T. N.; and Welling, M. 2017. Semi-Supervised Clas-
+sification with Graph Convolutional Networks. In 5th In-
+ternational Conference on Learning Representations, ICLR
+2017, Toulon, France, April 24-26, 2017, Conference Track
+Proceedings . OpenReview.net.
+Li, C.; Conejo, A. J.; Liu, P.; Omell, B. P.; Siirola, J. D.; and
+Grossmann, I. E. 2022a. Mixed-integer linear programming
+models and algorithms for generation and transmission ex-
+pansion planning of power systems. European Journal of
+Operational Research , 297(3): 1071–1082.
+Li, C.; Conejo, A. J.; Siirola, J. D.; and Grossmann, I. E.
+2022b. On representative day selection for capacity ex-
+pansion planning of power systems under extreme operat-
+ing conditions. International Journal of Electrical Power &
+Energy Systems , 137: 107697.
+Li, Q.; Han, Z.; and Wu, X.-M. 2018. Deeper Insights into
+Graph Convolutional Networks for Semi-Supervised Learn-
+ing. In Proceedings of the Thirty-Second AAAI Confer-
+ence on Artificial Intelligence and Thirtieth Innovative Ap-
+plications of Artificial Intelligence Conference and EighthAAAI Symposium on Educational Advances in Artificial In-
+telligence , AAAI’18/IAAI’18/EAAI’18. AAAI Press. ISBN
+978-1-57735-800-8.
+Liu, Y .; Sioshansi, R.; and Conejo, A. J. 2017. Hierarchi-
+cal clustering to find representative operating periods for
+capacity-expansion modeling. IEEE Transactions on Power
+Systems , 33(3): 3029–3039.
+Mallapragada, D. S.; Papageorgiou, D. J.; Venkatesh, A.;
+Lara, C. L.; and Grossmann, I. E. 2018. Impact of model
+resolution on scenario outcomes for electricity sector sys-
+tem expansion. Energy , 163: 1231–1244.
+Park, J.; Lee, M.; Chang, H. J.; Lee, K.; and Choi, J. Y . 2019.
+Symmetric Graph Convolutional Autoencoder for Unsuper-
+vised Graph Representation Learning. 2019 IEEE/CVF In-
+ternational Conference on Computer Vision (ICCV) , 6518–
+6527.
+Parsons, L.; Haque, E.; and Liu, H. 2004. Subspace Cluster-
+ing for High Dimensional Data: A Review. SIGKDD Explor.
+Newsl. , 6(1): 90–105.
+Reuther, A.; Kepner, J.; Byun, C.; Samsi, S.; Arcand, W.;
+Bestor, D.; Bergeron, B.; Gadepally, V .; Houle, M.; Hubbell,
+M.; et al. 2018. Interactive supercomputing on 40,000 cores
+for machine learning and data analysis. In 2018 IEEE High
+Performance extreme Computing Conference (HPEC) , 1–6.
+IEEE.
+Salha, G.; Hennequin, R.; and Vazirgiannis, M. 2019. Keep
+it simple: Graph autoencoders without graph convolutional
+networks. arXiv preprint arXiv:1910.00942 .
+Scott, I. J.; Carvalho, P. M.; Botterud, A.; and Silva, C. A.
+2019. Clustering representative days for power systems gen-
+eration expansion planning: Capturing the effects of vari-
+able renewables and energy storage. Applied Energy , 253:
+113603.
+Shuman, D.; Narang, S. K.; Frossard, P.; Ortega, A.; and
+Vandergheynst, P. 2012. The Emerging Field of Signal
+Processing on Graphs: Extending High-Dimensional Data
+Analysis to Networks and Other Irregular Domains. IEEE
+Signal Processing Magazine , 30.
+SI. 2022. Supplementary material available at:
+https://shorturl.at/bkHOU.
+Taubin, G. 1995. A Signal Processing Approach to Fair
+Surface Design. In Proceedings of the 22nd Annual Con-
+ference on Computer Graphics and Interactive Techniques ,
+SIGGRAPH ’95, 351–358. New York, NY , USA: Associa-
+tion for Computing Machinery. ISBN 0897917014.
+Teichgraeber, H.; and Brandt, A. R. 2019. Clustering meth-
+ods to find representative periods for the optimization of en-
+ergy systems: An initial framework and comparison. Ap-
+plied energy , 239: 1283–1293.
+UN-FCCC. 2015. Decision 1/CP. 21, Adoption of the Paris
+Agreement. In Report of the Conference of the Parties on Its
+Twenty-First Session, Held in Paris from , volume 30.
+Weiss, J.; and Hagerty, J. M. 2019. Achieving 80 %GHG
+Reduction in New England by 2050.
+