Incorporating spatial heterogeneity into landslide susceptibility assessment: a GWRF–SHAP new method

Landslide susceptibility assessment is essential for effective risk mitigation; however, conventional machine learning models often struggle to capture spatial heterogeneity and lack interpretability, resulting in limited generalization capability. To address these challenges, this research proposes an integrated method that combines Geographically Weighted Random Forest (GWRF) with Shapley Additive exPlanations (SHAP) for landslide susceptibility assessment. Focusing on a mountainous canyon region characterized by complex geological conditions and a vast spatial extent, RF and GWRF landslide susceptibility models were constructed from both global and subregional perspectives using watershed-based partitioning. Subsequently, SHAP was adopted to interpret model outputs and enhance interpretability and transparency. The results show that at both the global and subregional scales, the GWRF model consistently outperforms the conventional random forest model, with AUC values over 0.85. The predicted zones with high and very high landslide susceptibility are primarily spread along either side of river valleys, closely matching the observed landslide distribution. The DEM, Topographic Wetness Index (TWI), and distance to road were the dominant factors, with landslide susceptibility being positively correlated with TWI and negatively correlated with the other two factors. Overall, the GWRF–SHAP framework provides a promising way for landslide risk assessment in geologically complex areas.