Circumpolar Wall-to-Wall Detection of Retrogressive Thaw Slumps with Scarce Labels and Extreme Class Imbalance

Abstract

Retrogressive thaw slumps (RTS) are among the most dynamic ground deformations caused by permafrost thaw. Abrupt thaw processes such as RTS could roughly double the contribution of permafrost thaw to climate warming, and already threaten a third of Arctic infrastructure. Circumpolar RTS mapping has advanced quickly, yet released inventories together cover only part of the permafrost region. RTS occupy only 0.1-0.5% of image tiles and labels are costly, so a detector must be designed and evaluated as a rare-target problem. We therefore developed a deep learning system and mapped 20.68 M square kilometres of circumpolar permafrost using 41.6 M tiles of 4.8 m PlanetScope imagery. Training used 22,259 tiles with 2,609 RTS instances in 1,718 of them, and the model was tuned through 172 controlled experiments. A convolutional network matched every alternative tested, including multimodal fusion, foundation-model encoders and masked-autoencoder pretraining, each at far more compute for similar or worse results. Labelled data moved performance most, and a foundation model's advantage appeared only when labels were scarce. Tested once on held-out ecoregions at the real-world prevalence of 0.1-0.5%, the final model achieved an area under the precision-recall curve of 0.812-0.855, an object precision of 0.854 and F1 of 0.644. Deployed wall-to-wall, the model produced a circumpolar map of 19,068 high-confidence detections whose sampled precision ranges from 0.54 to 0.90 by size class. We present an evaluation protocol for rare targets, a controlled account of what governs accuracy under scarce labels, and the circumpolar RTS map the validated model produced.