2026 GeoAI Arctic Challenge

The GeoAI Arctic Challenge dataset is an instance segmentation benchmark for detecting and delineating retrogressive thaw slumps (RTS) in Arctic image chips. The dataset builds on Yang et al. (2023), which provided semantic segmentation masks labeling each pixel as RTS or non-RTS. For this challenge, those labels have been extended and reformatted so each RTS feature is represented as an individual instance.

Participants receive multimodal image data and train models to predict one mask for each RTS instance in the hidden test set. This instance-level formulation supports feature-level evaluation, including how well models separate and delineate individual RTS boundaries.

Training labels are provided in COCO instance segmentation format. Test labels remain hidden and are used by the official scorer.

Satellite Image (RGB)	Semantic Mask (Original)	Instance Mask (This Challenge)

Conversion from semantic RTS labels into instance-level masks. The challenge dataset uses connected-component instance labels so models can be evaluated at the feature level rather than only pixel-wise.

Geographic Coverage & Study Sites

The source data spans seven Arctic subregions, including:

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

Spatial coverage of the source Arctic RTS dataset. Credit: Li et al., 2025.

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

Public Release Contents

The downloadable files are hosted in the Hugging Face dataset repository. This page documents the release contents, formats, bands, statistics, and visual examples.

The dataset, starter scripts, and challenge documentation are available before the benchmark phase so participants can explore the data, train models, and prepare submissions. Benchmark submission and leaderboard-ranking access opens on August 15, 2026, when organizers add approved Hugging Face usernames to the submission portal.

The public package contains training images and labels, hidden-label test images, metadata, and starter tools:

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

Image Bands

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

Data layer	Source / feature type	Bands	Role in RTS mapping
RGB imagery	Maxar optical imagery	red, green, blue	Provides high-resolution visual context for exposed soil, vegetation disturbance, and RTS morphology.
Spectral features	Vegetation, water, and near-infrared features	ndvi, ndwi, nir	Helps distinguish thaw-related disturbance from vegetation, water, snow, and other surface conditions.
Terrain features	ArcticDEM-derived topographic features	relative_elevation, shaded_relief	Adds terrain structure that can improve boundary delineation.

The array band order is:

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 available in machine-readable form in metadata/band_names.json.

Key Dataset Statistics

Property	Description
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"}
Original Source	Yang et al. (2023), Remote Sensing of Environment

Label Conversion

Training annotations are stored in train/annotations/instances_train.json using 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.

Dataset Distributions

RTS Size Distribution

RTS Coverage Distribution

RTS Shape Analysis

Band Statistics

Sample Visualizations

Description	Satellite Image (RGB)	Instance Mask (RTS Features)
Single large RTS
Single small RTS
Multiple RTS
RTS near snow