Multi-geohazard susceptibility assessment and influencing factors in Zhejiang Province, China: a machine learning approach

Geohazards such as collapses, landslides, and debris flows result from complex interactions between human activities and environmental conditions. However, a quantitative understanding of their coupling mechanisms remains challenging. This study developed a machine learning-based classification framework​ for multi-geohazard susceptibility mapping (GSM) in Zhejiang Province, China, to address this gap. The study employed XGBoost, AdaBoost, and Random Forest to construct individual models for each geohazard, using a comprehensive set of geomorphologic, geological, environmental, hydrological, and anthropogenic factors. The XGBoost model achieved Area Under the Curve (AUC) values greater than 0.9 for all geohazards, and was selected as the optimal model for GSM. Results show: (1) Topographic position index (TPI) and distance to roads are the most influential factors, with dominant roles varying by geohazard—TPI primarily controls debris flows, while collapses are more driven by road proximity. (2) Anthropogenic factors account for 15.9%–33.8% of importance across geohazards. (3) The dependence plots and heatmap of interaction values reveal the impact of human–natural factor coupling mechanisms on geohazards. The study provides a quantitative and interpretable analysis of human–natural environment coupling, offering insights for risk management and spatial planning in densely populated coastal regions under climate change.