QCR-Underwater-Perception/reef-guidance-system

Dataset Card for Reef Guidance System

This dataset provides imagery used for training and evaluation of models in the Reef Guidance System. All imagery was collected by the Australian Institute of Marine Science using the ReefScan™ Transom Marine Monitoring System.

If you use this dataset in your work, please cite the associated paper: AI-driven dispensing of coral reseeding devices for broad-scale restoration of the Great Barrier Reef (citations provided at bottom of this page).

Acknowledgement of Country

We acknowledge the continuing Sea Country management and custodianship of the Great Barrier Reef by Aboriginal and Torres Strait Islander Traditional Owners whose rich cultures, heritage values, enduring connections and shared efforts protect the Reef for future generations.

The Queensland University of Technology (QUT) acknowledges the Turrbal and Yugara, as the First Nations owners of the lands where QUT now stands. We pay respects to their Elders, lores, customs and creation spirits. We recognise that these lands have always been places of teaching, research and learning. QUT acknowledges the important role Aboriginal and Torres Strait Islander people play within the QUT community.

Dataset Details

This dataset contains high resolution images of the seafloor. An example image is shown as follows:

This dataset currently contains 14,528 images of the seafloor, collected at depth ranges up to 10m. There is a combination of whole images and also a subset containing smaller image patches, as detailed in the following sections.

Description

This data was collected and labelled to train a model to visually determine suitable locations to dispense coral reseeding devices, as outlined in the accompanying manuscript.

We provide multiple subsets with different labelling schemes for this dataset. Different labelling regimes with differing levels of supervision support weakly and semi-supervised learning.

• Dataset Contact: Dr Scarlett Raine, QUT Centre for Robotics. Contact: [email protected]
• Funded By: This work was a collaboration between the QUT Centre for Robotics and the Australian Institute of Marine Science, as part of the Reef Restoration and Adaptation Program.
• License: cc-by-nc-sa-4.0
• Repository: GitHub
• Paper: Arxiv

To read more about the Reef Restoration and Adaptation Program: RRAP

To read more about the ReefScan™ Transom Marine Monitoring System: ReefScan™

Structure

We provide multiple labelling subsets for each location. The different subsets do not overlap i.e. you may train on Weak Labels or Patches and evaluate on the Deployment Sequences.

All images were collected on the Great Barrier Reef in Australia, with the specific sites in the table below.

The table below specifies the locations, depth ranges and labelling subsets for each of the 5 sites included in this dataset.

Site	Location	Depth Range	Deployment Sequences	Patches	Weak Labels
1	Combined Sites: Heron Island, Cairns, Moore Reef	1.4 - 10.4m	No	Yes, 2,191 patches	Yes, 586 images
2	Maureen's Cove, Whitsundays	1.9 - 6.4m	Yes, 1,000 images	No	Yes, 1,043 images
3	Black Island, Whitsundays	2.2 - 5.8m	Yes, 500 images	Yes, 1,944 patches	Yes, 582 images
4	Unsafe Passage, Whitsundays	1.4 - 8.3m	Yes, 1,000 images	Yes, 3,000 patches	Yes, 359 images
5	Heron Island	1.5 - 8.5m	Yes, 1,500 images	Yes, 3,000 patches	Yes, 600 images

Data Collection and Processing

All imagery collected by the Australian Institute of Marine Science using the ReefScan™ Transom Marine Monitoring System. The ReefScan™ is attached to the back of a boat or tender and is used to collect downward facing images of the seafloor. No post-processing or image enhancement techniques are applied to the images.

The patches are created by slicing the original images (5312 x 3040 pixels) into a grid of 5 rows and 9 columns, resulting in patches of size 758 x 760 pixels.

The whole images have associated metadata including:

• GPS location: Latitude, Longitude
• Depth: altitude or distance to seafloor
• Camera properties: focal length, exposure time, f-stop, ISO speed

Deployment Sequences are sequential frames to match the real coral dispensing scenario, labelled at the image-level as Deploy or No-Deploy. We provide HTML maps which visualise the GPS coordinates of the collected imagery. In these maps, red denotes a No-Deploy label, and green markers represent images labelled as Deploy.

Weak Labels are whole images that have been selected by our annotator to best represent the class i.e. the majority of the image matches the label. These images are labelled into Deploy, No-Deploy or Coral. The Coral class is also considered as a case in which we would not dispense a coral device. We include this extra class to facilitate a parallel task for mapping coral coverage.

Patches are smaller image crops which have been assigned a label. The patches are also labelled into Deploy, No-Deploy and Coral. We provide a training/test split for this label type.

Annotations

This dataset was labelled by:

Dr Amy Coppock, Marine Scientist, Australian Institute of Marine Science

We acknowledge and thank Amy for her significant contributions to this project.

Citation

If this dataset or the associated repository contributes to your research, we ask that you please cite the publication below.

BibTeX:

@article{raine2025ai,

title={AI-driven dispensing of coral reseeding devices for broad-scale restoration of the Great Barrier Reef},

author={Raine, Scarlett and Moshirian, Benjamin and Fischer, Tobias},

journal={arXiv preprint arXiv:2509.01019},

year={2025}

}

APA:

Raine, S., Moshirian, B., & Fischer, T. (2025). AI-driven dispensing of coral reseeding devices for broad-scale restoration of the Great Barrier Reef. arXiv preprint arXiv:2509.01019.

Dataset Card Authors

Dataset Contact: Dr Scarlett Raine, QUT Centre for Robotics, email: [email protected]

Dr Benjamin Moshirian, Australian Institute of Marine Science

Dr Tobias Fischer, QUT Centre for Robotics