Multi-level InSAR Coupling and Spatially Constrained Imbalanced Sampling for Refined Landslide Susceptibility Assessment

Machine learning-based landslide susceptibility assessment (LSA) faces critical challenges in landslides data scarcity for mountainous, negative sample uncertainty, sample label distortion and reliance on sample balance. Therefore, this study took the eastern Himalayan syntaxis (EHS) with extreme climate, high-steep terrain, and active structure as the study area, and proposed a novel semi-supervised imbalanced LSA framework with multi-level InSAR applications. First, improving the landslide inventory by SBAS-InSAR and high-resolution imagery. Second, selecting non-landslide samples through dual-constrained strategy combining Moran\u2019s index and frequency ratio (FR) thresholds, based on kernel density estimation (KDE). Third, obtaining initial landslide susceptibility indices (LSI) through random forest (RF). Fourth, excluding landslide samples from very low and low susceptibility zones, and reconstructed training set by selecting landslide samples (label=1) and non-landslide (label=initial LSI) samples from the same area with ratios ranging from 1:1 to 1:10. Finally, implementing semi-supervised imbalanced LSA using regression forest, and introduced SBAS-InSAR slope deformation rate to obtain final LSA result. Results demonstrate that the Moran-FR-based initial LSA model achieved superior performance (AUC = 0.998) over conventional LSA model. The optimal semi-supervised model (imbalance ratio = 1:5) yielded an AUC of 0.999 and successfully identified high-risk zones in the Sedongpu Gully after SBAS-InSAR slope deformation rate optimization. SHAP analysis revealed elevation, terrain relief, distance to road\/river, rainfall, NDVI and slope as dominant factors, and the role of strike slip faults with limited vertical activity in landslide development was relatively weak. Furthermore, NDVI displayed a hydrological negative feedback effect beyond critical vegetation thresholds. This study established a Moran-FR dual-constrained sampling method and a novel semi-supervised imbalanced LSA framework with multi-level InSAR applications, providing a new paradigm for LSA in EHS areas with high and steep terrain and active structure, and offering a transferable solution for the prevention and control of geological disasters similar to orogenic belts.