An intelligent thermokarst landslide detection method incorporating physics-driven multi-feature integration in cold–arid alpine regions

Global warming accelerates permafrost degradation in cold–arid alpine regions, resulting in frequent thermokarst landslides that pose serious threats to transport corridors and regional ecological security. In these environments, pronounced surface fragmentation, weak spectral contrast between landslide deposits and surrounding materials, and dynamic water–thermal disturbances lead to blurred boundaries and unstable feature expressions, limiting the effectiveness of single-modality optical or SAR-based detection methods. To address these challenges, this study proposes an intelligent thermokarst landslide detection method based on physics-driven multi-feature integration. Guided by the ‘water–heat–mechanics’ coupling mechanism, a multidimensional feature set integrating optical, SAR, DEM, spectral indices, and texture metrics is constructed to capture moisture-related variations, roughness contrasts, freeze–thaw disturbances, and topographic constraints. A lightweight TokenKAN-UNet is further developed by combining KANConv2D for nonlinear feature extraction, token-level global modelling, and multi-scale feature fusion, enabling precise boundary delineation with high efficiency. Experiments along the China–Pakistan Friendship Highway on the Pamir Plateau show that the proposed method achieves higher IoU, F1-score, and Recall than U-Net, U-Net++, and TransU-Net, while requiring only 15.18 M parameters and 31 ms/step inference. PDP analysis provides a descriptive interpretation of model responses in relation to the physical rationale for feature selection.