Enhancing geological hazard risk assessment through stacking ensemble learning

Geological hazards pose increasing threats to human society, highlighting the need for accurate risk assessment. This study develops a Stacking ensemble framework that integrates CatBoost, LightGBM, XGBoost, and Random Forest to assess geological hazard susceptibility and risk in Linxia Hui Autonomous Prefecture, Gansu Province, China. Multisource geospatial data – including topography, geology, climate, population, and buildings – were incorporated to capture complex hazard-driving mechanisms. Model interpretability was enhanced using SHapley Additive exPlanations (SHAP), which identified fault distance and elevation (DEM) as the dominant factors. The Stacking model achieved an accuracy of 85.45% and an AUC of 0.910, outperforming all individual models. Risk mapping revealed that high-risk zones for population loss are widely distributed across residential and rural areas, whereas high-risk zones for building loss are concentrated in urbanised centres. The proposed framework goes beyond susceptibility-only studies by systematically integrating ensemble learning, susceptibility assessment, interpretability analysis, and vulnerability-based risk assessment. These results provide a robust scientific basis for disaster prevention, landuse planning, and emergency management in the Yellow River Basin and similar high-risk regions.