The Cyber2A challenge website is the main information hub for the challenge overview, rules, dataset documentation, submission format, important dates, and updates:

https://cyber2a.github.io/challenge/

The Hugging Face Space is the submission and leaderboard portal:

https://huggingface.co/spaces/cyber2a/2026GeoAIArcticChallenge

Dataset Summary

The GeoAI Arctic Challenge dataset is an instance segmentation benchmark for detecting and delineating retrogressive thaw slumps (RTS) in Arctic image chips.

RTS are landslide-like permafrost landforms that occur when ice-rich permafrost thaws and collapses. Mapping RTS across Arctic regions can support permafrost degradation monitoring, environmental change detection, and research on climate-driven Arctic landscape transformation.

The dataset builds on the RTS dataset from Yang et al. (2023), which provided semantic segmentation masks labeling pixels as RTS or non-RTS. For this challenge release, those labels were extended and reformatted so each RTS feature is represented as an individual instance. Training annotations are provided in COCO instance segmentation format. Test labels are hidden and used for leaderboard evaluation.

Download

Download and unzip the released package:

unzip 2026_geoai_arctic_challenge_v1.0.zip

Expected top-level directory:

competition_release/

Package Contents

competition_release/
  README.md
  metadata/
    band_names.json
    sample_submission.json
    train_manifest.csv
    test_manifest.csv
  tools/
    coco_utils.py
    validate_submission.py
    evaluate_coco.py
    inspect_dataset.py
  examples/
    load_image_and_label.py
    make_sample_submission.py
    encode_predictions.py
  train/
    images/*.npz
    annotations/instances_train.json
  test/
    images/*.npz

Each .npz image file contains one array named image with shape H x W x 8 in HWC order.

Task

Participants train models to predict one mask for each RTS instance in the hidden-label test image chips.

Input	Output
Eight-band `.npz` image chip	One compressed COCO RLE mask per predicted RTS instance

There is one prediction category:

{"id": 1, "name": "rts", "supercategory": "landform"}

Dataset Statistics

Property	Value
Training images	756 image chips with public labels
Test images	138 image chips without public labels
Image array format	`.npz` files containing `image` arrays with shape `H x W x 8`
Annotation format	COCO instance segmentation JSON with compressed RLE masks
Task	RTS instance segmentation
Category	`{"id": 1, "name": "rts", "supercategory": "landform"}`

Image Bands

Each image chip contains eight co-registered channels combining optical imagery, spectral features, and topographic context.

Index	Band name	Description
0	`red`	Maxar red
1	`green`	Maxar green
2	`blue`	Maxar blue
3	`ndvi`	Normalized Difference Vegetation Index
4	`relative_elevation`	Relative elevation
5	`shaded_relief`	Shaded relief
6	`nir`	Planet near-infrared
7	`ndwi`	Normalized Difference Water Index

The same band list is provided in:

metadata/band_names.json

Geographic Coverage

The source data spans Arctic subregions in Canada and Russia, including:

Canada: Herschel Island, Horton Delta, Tuktoyaktuk peninsulas, Banks Island
Russia: Yamal and Gydan peninsulas, Lena River, Kolguev Island

The challenge release removes geospatial metadata from distributed image chips while preserving multimodal image information for modeling.

Labels

Training annotations are stored in:

train/annotations/instances_train.json

Annotations use COCO instance segmentation format. Instance masks are encoded as compressed COCO run-length encoding (RLE).

The label conversion rule is:

RTS foreground: finite source rts_label values greater than 0
Background: source rts_label == 0 or missing/no-label values
Instances: 8-connected components over the binary RTS foreground
Filtering: connected components smaller than 10 pixels are removed

This conversion is deterministic. If two RTS features touch in the source mask, connected-component labeling treats them as one instance.

Submission Format

Submissions use COCO results-format JSON with compressed RLE masks. See the full submission instructions on the Cyber2A website:

https://cyber2a.github.io/challenge/submission.html

Validate a submission before upload:

python tools/validate_submission.py --submission path/to/submission.json

The public validator checks JSON structure, image IDs, category IDs, score values, compressed RLE decodability, and mask sizes.

Each team may submit up to 2 times per day through the Hugging Face submission portal.

Intended Use

This dataset is intended for:

Participation in the 2026 GeoAI Arctic Challenge.
Research on RTS detection and delineation.
Benchmarking GeoAI, computer vision, remote sensing, and multimodal segmentation methods.
Educational use in AI, geospatial science, Earth science, and Arctic research contexts.

Data Usage and Redistribution

The dataset is provided for research and competition purposes within this challenge.

Redistribution of the dataset or derivatives outside the challenge requires permission from the original data owners and curators.

Participants may use external data, but external data usage must be documented in the method description or technical report.

Do not share labels, private data, submissions, or predictions across teams.

Challenge Links

Main challenge website: https://cyber2a.github.io/challenge/
Dataset documentation: https://cyber2a.github.io/challenge/dataset.html
Rules: https://cyber2a.github.io/challenge/rules.html
Submission format: https://cyber2a.github.io/challenge/submission.html
Resources and FAQ: https://cyber2a.github.io/challenge/resources.html
Submission portal: https://huggingface.co/spaces/cyber2a/2026GeoAIArcticChallenge

Citation

Works or publications using this dataset should cite the source RTS dataset paper and the related multimodal GeoAI RTS modeling paper.

Dataset Source

Yang, Yili, Brendan M. Rogers, Greg Fiske, Jennifer Watts, Stefano Potter, Tiffany Windholz, Andrew Mullen, Ingmar Nitze, and Susan M. Natali. "Mapping retrogressive thaw slumps using deep neural networks." Remote Sensing of Environment 288 (2023): 113495. https://doi.org/10.1016/j.rse.2023.113495

@article{yang2023mapping,
  title={Mapping retrogressive thaw slumps using deep neural networks},
  author={Yang, Yili and Rogers, Brendan M and Fiske, Greg and Watts, Jennifer and Potter, Stefano and Windholz, Tiffany and Mullen, Andrew and Nitze, Ingmar and Natali, Susan M},
  journal={Remote Sensing of Environment},
  volume={288},
  pages={113495},
  year={2023},
  publisher={Elsevier}
}

Related Multimodal GeoAI Model

Li, Wenwen, Chia-Yu Hsu, Sizhe Wang, Zhining Gu, Yili Yang, Brendan M. Rogers, and Anna Liljedahl. "A multi-scale vision transformer-based multimodal GeoAI model for mapping Arctic permafrost thaw." IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2025). https://doi.org/10.1109/JSTARS.2025.3564310

@article{li2025multi,
  title={A multi-scale vision transformer-based multimodal GeoAI model for mapping Arctic permafrost thaw},
  author={Li, Wenwen and Hsu, Chia-Yu and Wang, Sizhe and Gu, Zhining and Yang, Yili and Rogers, Brendan M and Liljedahl, Anna},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
  year={2025},
  publisher={IEEE}
}

Contact

For challenge and dataset support, contact:

For public technical questions that may help other teams, use the Discussions area on the Hugging Face submission portal.

Downloads last month: 28

Total file size:

561 MB