Landslide Susceptibility Evaluation Integrating Bayesian Model and Extremely Randomized Trees: A Case Study of Hanshan Subdistrict, Hanzhong, Shaanxi Province, China

Abstract Shaanxi Province spans three major geomorphological units with complex and fragile geological environments. Influenced by hazard-pregnant geological condition factors, the development of landslide geological disasters exhibits significant regional heterogeneity across different areas, thus urgently requiring geological disaster susceptibility evaluation based on sensitivity analysis of hazard-pregnant geological conditions to support subsequent geological disaster risk prevention and control in township regions. This study focuses on landslide disasters in Hanshan Subdistrict, Hanzhong City, Shaanxi Province. A Bayesian model based on hazard-pregnant geological conditions was constructed using Bayes' theorem, obtaining posterior probabilities of landslide development across classified intervals under various hazard-pregnant geological conditions, thereby identifying dominant development intervals for landslides under different predisposing factors. Considering the application of small-sample data streams in township areas, an Extreme Random Tree (ERT) model was further integrated to conduct landslide susceptibility evaluations using both the Bayesian model and the Bayesian-ERT coupled model. Results demonstrate that the Bayesian-ERT model achieves higher accuracy, classifying susceptibility zones as: extremely low (13.6%), low (20.1%), moderate (25.4%), high (22.8%), and extremely high (18.1%). The frequency ratios (1.719, 1.929) and AUC value (0.915) for high/extremely high susceptibility zones significantly surpass those of the standalone Bayesian model (1.323, 1.297, 0.827). This indicates that the Bayesian-ERT model not only precisely identifies high-risk areas in complex geological environments but also reduces misjudgment in low-probability regions, providing a high-precision and reliable technical tool for geological disaster prevention and control.