Slope unit-based comprehensive geohazard susceptibility assessment: SHAP interpretability and local InSAR deformation analysis

Abstract

Frequent geological hazards worldwide necessitate robust and comprehensive susceptibility models for effective risk management. This study focuses on four common hazards—landslides, rockfalls, unstable slopes, and debris flows—in the Three Gorges Reservoir Area (TGRA) and introduces a comprehensive geohazard susceptibility modeling approach. The method employs a Soft Voting ensemble learning model integrating Random Forest (RF) and Light Gradient Boosting Machine (LGBM). Slope unit-based susceptibility models were developed to calculate individual geohazard susceptibility indexes (GSIs) for each hazard type, which were then combined using a proposed weighted average method to conduct a comprehensive geohazard susceptibility assessment. The results indicate that the ensemble model outperformed base models across all hazard types, achieving area under the receiver operating characteristic (ROC) curve (AUC) of 0.752, 0.721, 0.813, and 0.841 for landslides, rockfalls, unstable slopes, and debris flows, respectively. The comprehensive geohazard susceptibility map (GSM) revealed that high-susceptibility zones concentrated along the Yangtze River, with recent geohazard events closely matching these areas. Frequency Ratio (FR) analysis showed a rapid increase in FR values with rising susceptibility levels, exceeding 1 in high and very high susceptibility zones. To analyze the deformation characteristics of typical high-GSI slope units in the reservoir area, InSAR time-series deformation analysis was conducted, revealing significant subsidence correlated with seasonal rainfall and reservoir water level fluctuations. Additionally, five key factors influencing geohazard occurrence were identified using SHapley Additive exPlanations (SHAP), offering localized interpretation for individual slope units. The findings overcome the limitations of single-hazard assessments and offer practical solutions for regional planning, infrastructure protection, and disaster mitigation. This research provides a valuable framework for policymakers and practitioners to enhance the resilience and ensure safety in geohazard-prone regions.