ellisbrown/BDML25SP_hw1_processed_data · Datasets at Hugging Face

text stringlengths 0 5.58k
burn areas. Burn incidents become indistinguishable within 2 days of burning as shown in Figure 5.
However, Sentinel-2 satellite imagery is freely available but on interval of 5-6 days. Infrared spectral
bands (SWIR) are available which are directly correlated with fire.
Daily monitoring of fire incidents is conducted at low spatial resolution using NASA satellite
instruments, specifically MODIS and VIIRS. The MODIS instrument captures data four times each
6

day, providing direct detection of fire events at a lower spatial resolution of 1 kilometer. This data is
typically available with a latency of 2 to 3 hours. Similarly, VIIRS collects data once a day, offering
direct detection of fires with an improved spatial resolution of 375 meters. Like MODIS, the data
from VIIRS is also available with a latency of 2 to 3 hours. These tools are critical for consistent and
timely monitoring of fire events over large areas.
Figure 4: Temporal sentinel imagery of burn area (May 2023)
Figure 5: Temporal planet imagery of burn area (May 2023)
A.1.2 Baseline Method
We conducted an analysis by comparing Planet and Sentinel imagery, along with active fire products,
to alert authorities and calculate the last burn index using Sentinel imagery. Figure 6 illustrates the
baseline method for detecting fire-affected areas using remote sensing data. The process begins
with two types of input data: active fire detection from MODIS/VIIRS and other remote sensing
datasets. These inputs are processed to generate visual representations of the affected regions, marked
by red circles. The processed images are then analyzed using various indices, including the Char
Index, Burn Area Index, Bare Soil Index, NBR (Normalized Burn Ratio), and others such as MIRBI
(Mid-Infrared Burn Index) and BSI (Burn Severity Index), to assess the extent and impact of the burn.
The outputs from these indices are subsequently used in a time-series change detection analysis to
monitor changes over time. The final result is visualized, indicating the spatial distribution of detected
changes (marked by red and blue triangles), which aids in identifying patterns and understanding the
impact of burning over time.
A.1.3 Preliminary Results
Combining MODIS, Sentinel, and Planet imagery significantly enhances the accuracy of fire detection,
building on methodologies proposed in prior studies. This integrated approach not only improves
detection precision but also facilitates the timely issuance of alerts to authorities, enabling prompt
action. In Figure 7, the first image, dated October 5, 2022, depicts the area before any burning
activity, showing a relatively uniform landscape. The second image, from October 16, 2022, shows
the aftermath of stubble burning, with darker patches clearly indicating fire-affected areas. The
third image, labeled ’Burn Area Mask’ and also dated October 16, 2022, precisely highlights the
locations impacted by the burning. The pink mask effectively outlines the burn areas, providing an
7

Figure 6: Workflow of baseline method using traditional remote sensing indices
accurate assessment of the extent of stubble burning. Similarly, the method successfully detected
stubble-burnt patches in another region, as shown in Figure 8. This visual analysis is crucial for
monitoring agricultural practices and assessing their environmental impact.
Figure 7: Masked burnt area occurred on 16thOct, 2022
Figure 8: Masked burnt area occurred on 9thMay, 2023
A.1.4 Limitation of baseline method and other techniques
Remote sensing (RS) spectral index-based method track the spectral differences between two images-
normal and burned. It is merely a difference of temporal changes in pixel that may be due to any
reason. Some of the RS based indices to detect burning are- MNDFI (Modified normalized difference
fire index), BAI (Burning Area Index), NBR (Normalized Burning Ratio).
Simple Models (CNN, RCNN etc.) can be used but not promising in case of limited training data. It
may also not detect temporal or positional relation. It may not deal with data from different sensors
of different resolution, for different local conditions. Whereas, Foundational models are trained on
diverse data from different sources and sensors.
8

A.2 Geospatial Foundation Model
Figure 9: The mask auto-encoder structure for pretraining Prithvi model on large scale multi-temporal
and multi-spectral satellite images [33].
Foundation models, trained on diverse datasets, are adapt at capturing temporal and spatial rela-
tionships, making them highly effective for complex tasks like stubble detection. Fine-tuning these
models with smaller, labeled datasets further enhances their accuracy. However, many farmers burn
stubble at night to avoid detection, making optical imagery alone insufficient. To address this, we need
to fine-tune the foundation model on a diverse range of data collected from multiple sensors. This data
should include optical and radar imagery, providing robust day/night coverage and mitigating issues
like cloud and noise interference. Such multi-modal data fusion cannot be effectively handled by
simple deep learning models. Instead, we require a foundation model trained on diverse datasets with
varying resolutions and sensor types. For this purpose, we selected the PRITHVI-100M geospatial
foundation model, which is specifically trained on three timestamps of Harmonized Landsat Sentinel
(HLS) data.
The PRITHVI-100M model represents a state-of-the-art approach to analyzing high-resolution
satellite imagery using advanced machine learning techniques. Built on the Vision Transformer (ViT)
architecture, it incorporates 3D patch embedding and 3D positional encoding to process multispectral
and temporal satellite data effectively. The model employs a self-supervised learning strategy based
on a masked auto-encoder (MAE). During training, multispectral images captured over various time
intervals and spectral bands are divided into smaller patches, which are flattened and processed by
the model. Its encoder-decoder structure generates a latent representation of the input, which is
used to reconstruct the original image. The training process is guided by a Mean Squared Error
(MSE) loss function to minimize reconstruction errors (Figure 9). The MAE learning strategy, which
involves masking certain patches during training, forces the model to learn underlying data patterns by
reconstructing the missing patches. This improves its ability to generalize and enhances its robustness
across applications like land cover classification, change detection, and environmental monitoring.
In this work, we fine-tune the PRITHVI-100M model on our dataset, using a Swin-B backbone
and a state-of-the-art U-Net regressor [49]. Unlike simpler models that struggle to integrate diverse
data types effectively, foundation models like PRITHVI-100M can leverage cross-modal learning to
capture nuanced relationships between different modalities. This capability significantly enhances
the accuracy and robustness of distinguishing stubble burning from other land disturbances.
PRITHVI-100M is particularly suitable for this task as it is trained on diverse earth observation data
across three timestamps, making it well-suited for change detection tasks. Additionally, it utilizes
six-band Harmonized Landsat Sentinel (HLS) data, including SWIR1 and SWIR2 bands. These
bands are highly effective in capturing the burning ratio, as demonstrated in the baseline method.
Overall, PRITHVI-100M represents a significant advancement in geospatial data analysis by combin-
ing ViT, 3D positional encoding, and MAE learning to deliver robust and scalable performance on
large-scale satellite imagery. After fine-tuning the model, the results will be validated against the
baseline method. However, challenges like detecting minor fires persist, underscoring the need for
further refinements.
9

DeepMyco - Dataset Generation for Dye
Mycoremediation
Danika Gupta

burn areas. Burn incidents become indistinguishable within 2 days of burning as shown in Figure 5.

However, Sentinel-2 satellite imagery is freely available but on interval of 5-6 days. Infrared spectral

bands (SWIR) are available which are directly correlated with fire.

Daily monitoring of fire incidents is conducted at low spatial resolution using NASA satellite

instruments, specifically MODIS and VIIRS. The MODIS instrument captures data four times each

6

day, providing direct detection of fire events at a lower spatial resolution of 1 kilometer. This data is

typically available with a latency of 2 to 3 hours. Similarly, VIIRS collects data once a day, offering

direct detection of fires with an improved spatial resolution of 375 meters. Like MODIS, the data

from VIIRS is also available with a latency of 2 to 3 hours. These tools are critical for consistent and

timely monitoring of fire events over large areas.

Figure 4: Temporal sentinel imagery of burn area (May 2023)

Figure 5: Temporal planet imagery of burn area (May 2023)

A.1.2 Baseline Method

We conducted an analysis by comparing Planet and Sentinel imagery, along with active fire products,

to alert authorities and calculate the last burn index using Sentinel imagery. Figure 6 illustrates the

baseline method for detecting fire-affected areas using remote sensing data. The process begins

with two types of input data: active fire detection from MODIS/VIIRS and other remote sensing

datasets. These inputs are processed to generate visual representations of the affected regions, marked

by red circles. The processed images are then analyzed using various indices, including the Char

Index, Burn Area Index, Bare Soil Index, NBR (Normalized Burn Ratio), and others such as MIRBI

(Mid-Infrared Burn Index) and BSI (Burn Severity Index), to assess the extent and impact of the burn.

The outputs from these indices are subsequently used in a time-series change detection analysis to

monitor changes over time. The final result is visualized, indicating the spatial distribution of detected

changes (marked by red and blue triangles), which aids in identifying patterns and understanding the

impact of burning over time.

A.1.3 Preliminary Results

Combining MODIS, Sentinel, and Planet imagery significantly enhances the accuracy of fire detection,

building on methodologies proposed in prior studies. This integrated approach not only improves

detection precision but also facilitates the timely issuance of alerts to authorities, enabling prompt

action. In Figure 7, the first image, dated October 5, 2022, depicts the area before any burning

activity, showing a relatively uniform landscape. The second image, from October 16, 2022, shows

the aftermath of stubble burning, with darker patches clearly indicating fire-affected areas. The

third image, labeled ’Burn Area Mask’ and also dated October 16, 2022, precisely highlights the

locations impacted by the burning. The pink mask effectively outlines the burn areas, providing an

7

Figure 6: Workflow of baseline method using traditional remote sensing indices

accurate assessment of the extent of stubble burning. Similarly, the method successfully detected

stubble-burnt patches in another region, as shown in Figure 8. This visual analysis is crucial for

monitoring agricultural practices and assessing their environmental impact.

Figure 7: Masked burnt area occurred on 16thOct, 2022

Figure 8: Masked burnt area occurred on 9thMay, 2023

A.1.4 Limitation of baseline method and other techniques

Remote sensing (RS) spectral index-based method track the spectral differences between two images-

normal and burned. It is merely a difference of temporal changes in pixel that may be due to any

reason. Some of the RS based indices to detect burning are- MNDFI (Modified normalized difference

fire index), BAI (Burning Area Index), NBR (Normalized Burning Ratio).

Simple Models (CNN, RCNN etc.) can be used but not promising in case of limited training data. It

may also not detect temporal or positional relation. It may not deal with data from different sensors

of different resolution, for different local conditions. Whereas, Foundational models are trained on

diverse data from different sources and sensors.

8

A.2 Geospatial Foundation Model

Figure 9: The mask auto-encoder structure for pretraining Prithvi model on large scale multi-temporal

and multi-spectral satellite images [33].

Foundation models, trained on diverse datasets, are adapt at capturing temporal and spatial rela-

tionships, making them highly effective for complex tasks like stubble detection. Fine-tuning these

models with smaller, labeled datasets further enhances their accuracy. However, many farmers burn

stubble at night to avoid detection, making optical imagery alone insufficient. To address this, we need

to fine-tune the foundation model on a diverse range of data collected from multiple sensors. This data

should include optical and radar imagery, providing robust day/night coverage and mitigating issues

like cloud and noise interference. Such multi-modal data fusion cannot be effectively handled by

simple deep learning models. Instead, we require a foundation model trained on diverse datasets with

varying resolutions and sensor types. For this purpose, we selected the PRITHVI-100M geospatial

foundation model, which is specifically trained on three timestamps of Harmonized Landsat Sentinel

(HLS) data.

The PRITHVI-100M model represents a state-of-the-art approach to analyzing high-resolution

satellite imagery using advanced machine learning techniques. Built on the Vision Transformer (ViT)

architecture, it incorporates 3D patch embedding and 3D positional encoding to process multispectral

and temporal satellite data effectively. The model employs a self-supervised learning strategy based

on a masked auto-encoder (MAE). During training, multispectral images captured over various time

intervals and spectral bands are divided into smaller patches, which are flattened and processed by

the model. Its encoder-decoder structure generates a latent representation of the input, which is

used to reconstruct the original image. The training process is guided by a Mean Squared Error

(MSE) loss function to minimize reconstruction errors (Figure 9). The MAE learning strategy, which

involves masking certain patches during training, forces the model to learn underlying data patterns by

reconstructing the missing patches. This improves its ability to generalize and enhances its robustness

across applications like land cover classification, change detection, and environmental monitoring.

In this work, we fine-tune the PRITHVI-100M model on our dataset, using a Swin-B backbone

and a state-of-the-art U-Net regressor [49]. Unlike simpler models that struggle to integrate diverse

data types effectively, foundation models like PRITHVI-100M can leverage cross-modal learning to

capture nuanced relationships between different modalities. This capability significantly enhances

the accuracy and robustness of distinguishing stubble burning from other land disturbances.

PRITHVI-100M is particularly suitable for this task as it is trained on diverse earth observation data

across three timestamps, making it well-suited for change detection tasks. Additionally, it utilizes

six-band Harmonized Landsat Sentinel (HLS) data, including SWIR1 and SWIR2 bands. These

bands are highly effective in capturing the burning ratio, as demonstrated in the baseline method.

Overall, PRITHVI-100M represents a significant advancement in geospatial data analysis by combin-

ing ViT, 3D positional encoding, and MAE learning to deliver robust and scalable performance on

large-scale satellite imagery. After fine-tuning the model, the results will be validated against the

baseline method. However, challenges like detecting minor fires persist, underscoring the need for

further refinements.

9

DeepMyco - Dataset Generation for Dye

Mycoremediation

Danika Gupta