Improving landslide susceptibility mapping with skeleton elastic storage coefficient and spatial learning models

Landslides pose substantial hazards in karst regions such as Bijie. However, the conventional subsidence metrics fail to represent the hydro-mechanical processes triggering slope failure. To address this limitation, the Skeleton Elastic Storage Coefficient (SKE) is proposed as a physically interpretable indicator. SKE was derived by integrating SBAS-InSAR deformation with GRACE-FO/GLDAS water storage data. To evaluate the performance of the SKE, a systematic comparison of 24 model-feature-sampling combinations was conducted, including traditional machine learning (SVM, RF, XGBoost) and deep learning models (MLP, LSTM, Transformer), two feature sets (SKE vs. subsidence), and two negative sampling schemes. The F1-score and AUC were used for evaluation. The results show that the SKE combinations consistently outperformed the subsidence metrics combinations, increasing the best F1-score by 4.50% and AUC by 3.61%. The best performance was achieved by the combination of Transformer with SKE and terrain-factor-filtered negatives (F1-score = 0.9498; AUC = 0.9905), reflecting its ability to capture long-range spatial dependencies. High-susceptibility zones concentrate along steep slopes and river corridors in Nayong, Hezhang, and Qixingguan. By linking surface deformation to subsurface hydrogeological processes, the SKE provides a more sensitive and interpretable basis for landslide susceptibility mapping than traditional deformation-only approaches.