Comparative Landslide Susceptibility Mapping in Longchuan, Guangdong Province, China, Using Explainable Machine Learning

Landslide susceptibility assessment is essential for disaster-risk reduction, land-use regulation, and territorial spatial planning in mountainous and hilly regions. However, the practical application of machine learning-based susceptibility models is often limited by the trade-off between predictive accuracy and model interpretability, as well as the instability of factor importance across different algorithms and study areas. Taking Longchuan County in northeastern Guangdong Province, China, as a case study, this research develops a comparative explainable machine learning framework to evaluate landslide susceptibility and examine the cross-model stability of SHAP-based factor attribution under local geo-environmental conditions. Fifteen conditioning factors were initially derived from multi-source geological, topographic, hydrological, environmental, and anthropogenic datasets. After multicollinearity screening using Pearson correlation analysis, twelve key factors were retained for model construction. A total of 363 historical landslide points and an equal number of non-landslide samples were divided into training and testing datasets using a stratified 70:30 sampling strategy. Eight machine learning models were optimized through grid-search parameter tuning and then comparatively evaluated. The results show that all models achieved strong predictive performance, with test-set AUC values exceeding 0.938. Among them, the Gradient Boosting Decision Tree model performed best, with an AUC of 0.9520 and the most stable control of overfitting, followed closely by CatBoost with an AUC of 0.9512. SHAP-based interpretation further revealed that the normalized difference water index, relief, and distance to rivers were the dominant factors controlling landslide susceptibility in the study area, with the normalized difference water index serving as a key explanatory factor across models. The proposed framework improves the transparency and reliability of landslide susceptibility assessment and provides a methodological reference for localized, explainable machine learning applications in geohazard risk management.