Automated detection of retrogressive thaw slumps in high-resolution optical imagery using a knowledge-distilled Transformer

Climate warming is accelerating Arctic permafrost degradation and intensifying retrogressive thaw slump (RTS) activity. To address the limited availability of RTS samples, large target-scale variation and unstable bounding-box localization in complex high-resolution optical remote-sensing imagery, this study proposes an improved DINO-based detection framework, SM-DINO. The framework uses cross-architecture feature-level knowledge distillation to transfer high-level semantic information from a ViT-S/16 teacher model pre-trained on a large remote-sensing data set to a ResNet-50 backbone. It further introduces a multi-scale dilated feature aggregation (MDFA) module and a geometry-modulated focal IoU (GMF-IoU) loss. Experiments on a high-resolution RGB imagery data set from Banks Island, Canada, show that SM-DINO achieves an mAP0.5:0.95 of 86.6% and mAP0.5 of 98.0%. These values are modestly higher than those of the original DINO baseline. The results suggest that remote-sensing knowledge transfer, multi-scale feature modelling and geometry-aware regression optimization can support RTS candidate-target detection in complex polar environments, while the method remains most suitable for server-side offline mapping rather than real-time deployment.