A comparative study on spatial patterns and dominant factors of landslide susceptibility under geomorphological differentiation

Spatial heterogeneity poses a significant challenge to large-scale landslide susceptibility assessment, often leading to generalized models that overlook local variations. Focusing on southeastern China, a region with complex terrain, this study integrates a geomorphological zoning strategy with the XGBoost algorithm to enhance prediction accuracy. The study area was partitioned into two distinct sub-regions based on environmental disparities, utilizing 14 conditioning factors. The comparative analysis reveals that the sub-region II model achieved superior performance with an AUC of 97.44%, significantly surpassing both the sub-region I model (85.86%) and the overall model (89.74%). This provides compelling evidence that the zoning strategy outperforms global modeling by eschewing the homogenization of complex terrain, thus precisely delineating local disaster mechanisms. Furthermore, SHAP analysis elucidated that dominant drivers are divergent across sub-regions, driven by unique local geographical contexts. Consequently, coupling regional partitioning with interpretable machine learning effectively captures local disaster mechanisms, providing a robust and precise scientific basis for disaster prevention and risk management.