2026 GeoAI Arctic Challenge

This page collects starter materials, learning resources, contacts, citations, and frequently asked questions.

Public challenge materials, including the dataset, starter tools, and documentation, are available before the benchmark phase so teams can train models and prepare submissions. Benchmark submissions and leaderboard rankings open on August 15, 2026, after approved Hugging Face usernames are added to the submission portal.

Quick Links

📘 Starter Materials

Helper scripts to load data, inspect labels, validate submissions, and encode masks.

Check the submission page or download the released files from the Hugging Face dataset repository.

🌍 Dataset

Data description, formats, bands, statistics, and visualization gallery.

View dataset

🧭 Rules

Step-by-step guide to register, submit, and score.

How to participate and submit

🏆 Help

Use the Hugging Face discussion area for public technical questions, or contact the organizers by email.

Contact

• Competition support: wenwen@asu.edu, chiayuhsu@asu.edu
• Technical Q&A: use the Hugging Face discussion area if available for public questions that may help other teams.

Starter Materials

The competition files include helper scripts for:

• Loading image chips and labels.
• Inspecting dataset metadata.
• Creating a sample submission.
• Encoding binary masks or polygons as compressed COCO RLE.
• Validating submission JSON files.
• Running local evaluation against public labels or a validation split.

Tutorials and Learning Materials

• PyTorch instance segmentation tutorial: https://docs.pytorch.org/tutorials/intermediate/torchvision_tutorial.html
• Detectron2 getting started guide: https://detectron2.readthedocs.io/en/stable/tutorials/getting_started.html
• COCO mask utilities: https://github.com/cocodataset/cocoapi
• NumPy .npz loading: https://numpy.org/doc/stable/reference/generated/numpy.load.html
• Hugging Face: https://huggingface.co

References

The GeoAI Arctic Challenge builds upon an RTS dataset and related multimodal GeoAI research. Participants should cite the following works when using the dataset or baseline model.

Dataset Source

Yang et al. (2023) developed a deep learning approach for mapping RTS from satellite imagery. This dataset forms the foundation of the competition.

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 et al. (2025) present a multi-scale vision transformer-based multimodal GeoAI model that integrates multispectral imagery and topographic features for RTS mapping in Arctic permafrost regions.

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}
}

FAQ

Where do I download the data?

Download the public release package from the Hugging Face dataset repository. You can review the dataset, train models, and prepare submissions before the benchmark phase opens. To submit predictions and view leaderboard rankings starting August 15, 2026, create a Hugging Face account, complete the registration form with each team member’s exact Hugging Face username, and wait for organizer approval.

What is the submission format?

Submissions use COCO results-format JSON with compressed RLE masks. See the Submission information page for the full schema.

Can I use external data?

Yes, but you must document all external data usage in your method description or technical report.

Can a team submit from multiple accounts?

No. Each team should register once. Starting August 15, 2026, all approved Hugging Face usernames for that team share the same team submission history, 2-submission daily limit, and leaderboard entry.

Are empty submissions allowed?

Yes. An empty JSON list ([]) is valid and means no RTS instances were predicted.

Do I need to submit code?

All teams must submit valid predictions. Top-performing teams must submit code or a reproducibility package for verification to be eligible for awards.