Landslide Hazard Zonation Driven by Multi-Rainfall Scenarios Based on the Optimal XGBoost Model—A Case Study of Yongren County, Yunnan Province, China

To address the limitations of low model accuracy and single-scenario settings in traditional rainfall-induced landslide hazard assessments, this study focuses on Yongren County, Yunnan Province—a region where landslides pose significant threats to sustainable socio-economic development and infrastructure resilience. Eight controlling factors—lithology, slope, terrain relief, distances to faults, rivers, and roads, vegetation coverage, and elevation—were used to build a landslide susceptibility index system. Three internationally recognized machine learning models, Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost), were applied for comparison. The XGBoost model was further coupled with rainfall scenario analysis, simulating three rainfall scenarios—normal, 10-year, and 20-year return periods—to form a framework integrating “high-precision susceptibility prediction–multi-scenario rainfall driving–dynamic hazard assessment.” Results show that XGBoost achieved the highest accuracy and stability, with AUC and overall accuracy exceeding those of RF and SVM, supporting high-precision multi-scenario simulations. High-hazard zones expanded from road-disturbed areas under normal rainfall to riverbanks under 10-year rainfall and to fault-fracture and road–river interaction zones under 20-year rainfall. This study provides a transferable framework for sustainable landslide risk management, enabling precision prevention, optimizing resource allocation for disaster risk reduction, and supporting evidence-based policy-making for sustainable development and climate adaptation in similar geological settings.