GeoAI Arctic Mapping Challenge Dataset

The GeoAI Arctic Mapping Challenge dataset builds upon Yang et al. (2023) and focuses on detecting and mapping retrogressive thaw slumps (RTS)—landscape disturbances caused by permafrost thaw. For this competition, we extend and reformat the dataset into an instance segmentation benchmark, enabling participants to train models that can better delineate individual RTS features across diverse Arctic regions.

Why it matters: RTS are sensitive indicators of permafrost thaw, which releases greenhouse gases and alters Arctic landscapes. By leveraging AI, we aim to accelerate RTS detection and improve understanding of climate-driven change.

Geographic Coverage & Study Sites

The dataset spans 7 Arctic subregions, including:

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

Figure 1. Spatial coverage of the GeoAI Arctic RTS dataset. The dataset includes 7 Arctic subregions across Canada and Russia, representing diverse geomorphic and climatic conditions (Li et al., 2025).

Data Sources & Multimodal Inputs

This dataset integrates multi-source satellite and geospatial data:

Data Type	Source	Resolution	Band Names	Purpose in Task
RGB Imagery	Maxar	4 m	maxarR, maxarG, maxarB	High-resolution base imagery for visual recognition
Multi-spectral	Sentinel-2	10 m	NDVI, NDWI, NIR	Vegetation & water indices for spectral feature learning
Elevation	ArcticDEM	2 m	relative elevation, shaded relief	Topographic context to improve RTS boundary detection

Annotation Strategy & Task Setup

Originally, Yang et al. (2023) provided semantic segmentation masks - binary labels indicating RTS vs. non-RTS regions.

For this challenge, we converted the dataset into instance segmentation format so each RTS feature is labeled individually.

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

Figure 2. Conversion from semantic to instance segmentation masks. Original semantic masks from Yang et al. (2023) labeled RTS vs. non-RTS, while the challenge dataset uses instance-level masks, enabling finer-grained evaluation and model learning.

Key Dataset Statistics

Property	Description
Total Regions	7 Arctic subregions
Total Images	756 train + 138 test
Total RTS Instances	2,110
Imagery Resolution	Maxar 4 m, Sentinel-2 10 m, ArcticDEM 2 m
Spectral Bands	RGB + NDVI, NDWI, NIR + DEM
Task	Instance segmentation
Labels	Per-instance RTS masks
File Formats	`.npz` images + `.json` COCO-style annotations
Original Source	*Yang et al.* (2023)*, Remote Sensing of Environment*