Abstract Landslide disasters cause severe loss of life and property worldwide, with Fuyuan County in Yunnan Province, Southwest China, being a representative example. Accurate landslide susceptibility mapping (LSM) is essential for effective disaster risk reduction. In this study, we applied four machine learning models, including ensemble models—Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM)—and Support Vector Machine (SVM), to model landslide susceptibility in Fuyuan County. Their performance was compared with traditional methods, including the Analytical Hierarchy Process (AHP) and Information Value (IV). Among all models, RF achieved the highest predictive accuracy, with a test AUC of 0.8420. XGBoost and LightGBM exhibited better robustness and generalization, as indicated by similar Kappa values (0.516 and 0.495) and RMSE (0.424 and 0.429). SVM and IV showed moderate performance, while AHP produced inferior results. Feature importance analysis identified lithology, mining density, and proximity to rivers as dominant factors, highlighting the combined effects of natural conditions and human activities. High-susceptibility zones were mainly distributed in the central, southwestern, and northeastern parts of Fuyuan County, consistent with historical records. This study demonstrates the advantages of machine learning models and provides practical insights and a replicable framework for disaster early warning and prevention applicable not only in Southwest China but also in other landslide-prone regions worldwide.
Cui et al. (Mon,) studied this question.