Physics-informed and interpretable assessment of seismic landslide susceptibility: a mechanistic-learning approach coupling PINN and SHAP

Earthquake-triggered landslides pose serious threats in tectonically active mountain regions, where reliable susceptibility assessment is needed for both pre-earthquake preparedness and post-earthquake rapid response. However, landslide susceptibility modelling remains challenged by uncertain non-landslide samples, limited physical interpretability, and difficulty in comparing susceptibility patterns before and after seismic disturbance. Taking the 2022 Luding earthquake as a case study, this study proposes an integrated framework combining Isolation Forest (IF), a physics-informed neural network (PINN), and SHapley Additive exPlanations (SHAP). IF was used to refine pseudo-negative samples, while a factor-of-safety-based consistency penalty was incorporated into the PINN loss function to encourage predictions to remain consistent with slope-stability logic. SHAP was used to interpret scenario-specific model-attribution patterns and guide feature optimization. The framework achieved competitive predictive performance under both pre- and post-earthquake conditions. SHAP-guided optimization identified a compact 13-factor subset that balanced predictive performance, model parsimony, and interpretability. Hydro-geomorphological predictors contributed more strongly before the earthquake, whereas seismic-tectonic predictors contributed more strongly after the earthquake. The proposed framework provides a physics-guided and interpretable strategy for landslide susceptibility assessment, supporting pre-earthquake planning and post-earthquake emergency prioritization.