2026 GeoAI Arctic Challenge

Registration and Setup

Who Can Participate

The competition is open to everyone, including students, academics, and industry professionals. Participation is organized at the team level. Each team may include 1 to 5 members and must designate one Team Leader to manage registration, Codabench access, and submissions. We strongly encourage and will give additional credit to teams composed of individuals with diverse backgrounds (e.g., Earth science + AI).

How to Register

  1. Form your team and select a Team Leader.
  2. Complete the team registration form.
  3. Provide the following information:
    • Team name
    • Team Leader’s name and email
    • Names and affiliations of all team members
    • Team Leader’s Codabench username (if already available)

The organizers will review all registrations and verify team eligibility.

Codabench Access and Resources

After submitting the registration form:

  1. The Team Leader should create a Codabench account at https://www.codabench.org.
  2. Visit the competition page: GeoAI Arctic Challenge.
  3. Click Join Competition.
  4. Wait for organizer approval.

Access requests are reviewed manually. Approved teams will receive access to the competition resources and leaderboard.

Once approved, the Team Leader can:

  1. Access the competition files and documentation.
  2. Download the public dataset and starter materials.
  3. Follow the setup instructions provided in the repository or competition page.
  4. Submit predictions through Codabench.

Participation Rules

  • Each team may register only once.
  • Each team must use a single Codabench account managed by the Team Leader.
  • Multiple Codabench accounts representing the same team are not permitted.
  • Teams found circumventing participation limits through multiple accounts may be disqualified at the organizers’ discretion.
  • Team membership should remain unchanged after registration unless explicitly approved by the organizers.
  • Codabench will enforce platform-level competition settings, including maximum submissions per day, maximum upload size, public/private leaderboard behavior, final ranking metric, and displayed leaderboard columns.
  • Code for the top-performing models must be submitted for verification to be eligible for awards.

Download Competition Files

Once registered and approved, download the public release package from the Files section on Codabench. You can also visit the Dataset Page for detailed information about the data, annotations, and visualizations.

Release package contents:

  • 756 training image chips and 138 test image chips.
  • Eight-band .npz image arrays with RGB, spectral indices, elevation, shaded relief, and NIR bands.
  • COCO-style instance segmentation annotations for the training split.
  • Hidden test labels used by the official scorer.
  • Helper scripts for loading images, validating submissions, encoding masks, and local evaluation.

Local Setup

Install the helper dependencies in your Python environment:

pip install -r requirements.txt

Inspect the release package:

python tools/inspect_dataset.py

Load one training image and decode one training instance mask:

python examples/load_image_and_label.py

Data Format

The release is organized into training data, test data, metadata, tools, and examples:

competition_release/
  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
Item Format Description
Imagery .npz One array named image with shape H x W x 8
Training labels .json COCO instance segmentation labels with compressed RLE masks
Manifests .csv Public IDs, integer image IDs, filenames, widths, and heights
Sample submission .json Example COCO results-format submission

The image bands are stored in HWC order: red, green, blue, ndvi, relative_elevation, shaded_relief, nir, and ndwi.

Task Description

Your task is to predict per-instance RTS masks for each test image chip.

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"}

Submission Format

Submit a COCO results JSON list. Each prediction should include:

{
  "image_id": 1,
  "category_id": 1,
  "segmentation": {"size": [159, 289], "counts": "01nk\\1"},
  "score": 0.95
}

Submission rules:

  • image_id must be an integer from metadata/test_manifest.csv.
  • category_id must be 1.
  • segmentation must be a compressed COCO RLE mask matching the image height and width.
  • score must be numeric and in [0, 1].
  • Empty submissions ([]) are valid and mean no predicted RTS instances.
  • Images with no predictions are valid.
  • Multiple predictions per image are valid.

If your model outputs binary masks, encode each mask as compressed RLE. If your model outputs polygons as x1, y1, x2, y2, ..., convert each polygon to compressed RLE before submission. Helper functions are provided in tools/coco_utils.py.

Create a valid sample submission:

python examples/make_sample_submission.py

Encode binary masks or polygon coordinates as compressed COCO RLE:

python examples/encode_predictions.py

Validate a submission before upload:

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

Malformed submissions should be fixed before upload. The public validator checks JSON structure, image IDs, category IDs, scores, compressed RLE decodability, and mask sizes.

Evaluation Metrics

Submissions are ranked on Codabench using COCO mask AP/AR (segm) with RTS-specific settings. The primary ranking metric is:

  • AP @[IoU=0.50:0.95, area=all, maxDets=10]

Additional displayed metrics may include:

Metric Definition
AP50 Average Precision at IoU 0.50
AP75 Average Precision at IoU 0.75
AP_small Average Precision for objects with area [0, 300] pixels
AP_medium Average Precision for objects with area [300, 2000] pixels
AP_large Average Precision for objects with area [2000, 1e10] pixels
AR_all Average Recall across all object sizes

Official settings:

  • maxDets: [1, 5, 10]
  • areaRngLbl: ["all", "small", "medium", "large"]
  • all: [0, 1e10] pixels
  • small: [0, 300] pixels
  • medium: [300, 2000] pixels
  • large: [2000, 1e10] pixels

The official metric uses the top-scoring detections per image according to maxDets=10. You may submit more than 10 predictions for an image, but predictions beyond the top 10 by score do not improve the official metric.

Evaluate predictions against the public training labels or your own COCO-format validation split:

python tools/evaluate_coco.py \
  --ground-truth train/annotations/instances_train.json \
  --submission path/to/train_predictions.json

Submission Instructions

  1. Train your model locally.
  2. Generate predictions for the test set.
  3. Encode each predicted RTS instance as compressed COCO RLE.
  4. Build one submission.json COCO results list.
  5. Validate the JSON with tools/validate_submission.py.
  6. Upload the validated submission through Codabench.
  7. Check your score on the leaderboard.

Required Materials

All teams must submit a valid COCO results-format submission.json through Codabench. In addition to leaderboard submissions, top-performing teams will be invited to contribute to an outcome publication and may be asked to provide the following materials.

Technical Report

Teams aiming for awards and paper inclusion must submit a short technical report (2-4 pages) covering:

  • Team and affiliations: members, institution, country.
  • Model architecture: CNN, Transformer, hybrid, or other approach.
  • Training strategy: losses, augmentations, optimizers, batch size, and hardware.
  • Special techniques: domain adaptation, self-training, ensembling, band selection, or other methods.
  • Results and ablations: validation metrics, ablation tables, and sample outputs.
  • Insights: challenges faced, lessons learned, and potential future improvements.
  • References to related works.

A report template will be provided to keep submissions consistent.

Code and Reproducibility Package

  • Final inference code or container to verify results.
  • Instructions for reproducing predictions on the test set.
  • Training scripts are encouraged but optional.

Important Dates

Event Date
Competition Opens June 1, 2026
Dataset Released June 1, 2026
Benchmark Phase Starts July 15, 2026
Final Submission December 31, 2026
Winners Announced February 1, 2027

Awards

Prizes will be awarded to the top-performing teams based on the final leaderboard standings.

  • 1st Place: $1,000 USD
  • 2nd Place: $500 USD
  • 3rd Place: $200 USD

Support and Leaderboard

Use the Codabench discussion forum for public challenge questions.

The leaderboard will be updated regularly with the latest submissions.