Regional dependence in earthquake-induced landslide susceptibility modeling: insights from Jiuzhaigou and Luding, China

Earthquake-induced landslide (EIL) susceptibility assessment in tectonically active mountainous regions remains challenging because the relationships among seismic forcing, geo-environmental conditions, and landslide occurrence vary across regions. This study develops an interpretable machine-learning framework to compare model performance and learned decision patterns across two earthquake-affected regions in southwestern China: Jiuzhaigou (2017 earthquake) and Luding (2022 earthquake). Four machine-learning models—random forest, support vector machine, extreme gradient boosting, and multilayer perceptron—are evaluated alongside four ensemble strategies using slope-unit susceptibility mapping (9,400 units in Jiuzhaigou and 9,712 in Luding) and multisource conditioning factors. Shapley Additive Explanations (SHAP) quantify factor contributions and reveal regional heterogeneity in model decision patterns. The results reveal regional differences in predictive performance and decision patterns. Ensemble strategies generally achieved competitive performance relative to individual models in both regions. All modeling strategies showed higher cross-validated AUC values in Jiuzhaigou than in Luding, suggesting that regional characteristics influence model predictability. SHAP further shows that peak ground acceleration makes the largest contribution in Jiuzhaigou, whereas mean annual rainfall dominates in Luding. Despite using the same conditioning-factor system and modeling framework, the models revealed distinct regional factor-contribution patterns and response relationships, highlighting the importance of regional validation and interpretation in EIL susceptibility modeling.