Arctic-specific AI model beats general Earth observation benchmarks by 15%

A new Arctic-domain remote sensing foundation model (RSFM) has been introduced by researchers at the University of Connecticut and partners. The team tackled the challenge of analyzing very-high-spatial-resolution (VHSR) satellite imagery in the Arctic by first curating a diverse training dataset from 267 terabytes of Vantor imagery. They used affinity-propagation clustering on spectral and acquisition-metadata descriptors to select approximately 3 million non-redundant chips, avoiding oversampling of repetitive or low-information areas. This diversity-aware curation fed a ViT-Large encoder pretrained with a masked autoencoder (MAE) reconstruction objective, yielding domain-specific transformer weights tailored for Arctic features.

Evaluated across four hand-labeled datasets—infrastructure, IWP, RTS, and TCNs—the model achieved foreground mean F1 scores of 0.87, 0.72, 0.93, and 0.87 respectively, representing a 5-8 percentage point improvement over an ImageNet-initialized ViT-Large baseline. More significantly, it outperformed the general-purpose Earth observation model Prithvi-EO-2.0 by at least 15% on all tasks. These results demonstrate that optimizing the pretraining data distribution at a regional scale, while keeping the architecture fixed, can produce highly transferable representations for fine-scale Arctic mapping—a critical capability for monitoring climate-driven changes in polar regions.