Abstract
Abstract
The intensification of retrogressive thaw slumps (RTSs) across permafrost regions in reaction to anthropogenic climate change presents critical challenges to infrastructure and ecosystems. This begs the question of quantifying disaster triggers and regional-scale spatial susceptibility of RTS. Presented here, an innovative multi-methodological framework combining multivariate environmental factor analysis, deep learning-enhanced feature extraction, and ensemble modeling to evaluate RTS susceptibility. It was performed in permafrost regions of Hoh Xil, Qinghai-Tibet Plateau, where unique geography engendered a proliferation of RTSs. Analysis revealed hydrothermal dynamics as the primary trigger of RTS activation, particularly when encountering extreme rainfall and anomalous freeze-thaw degree days. Ground ice content and fine-grained soil composition constitute essential material prerequisites for failure development. Besides, negative sampling via clustering with geological similarity constraints effectively mitigated sample imbalance and enhanced model generalizability. The integration of convolutional neural networks (CNNs) effectively captured terrain-induced signatures associated with RTS initiation, enabling the generation of operationally susceptible maps with enhanced spatial precision. The heterogeneous ensemble modeling comparison demonstrated that the Boosting-integrated framework achieved optimal predictive performance, yielding an accuracy of 0.9853, significantly outperforming base models. The integrated deep learning and geostatistics approach decodes complex RTSs triggers and delivers scalable solutions for climate-adaptive engineering planning in permafrost regions.