Stacking ensemble learning-driven risk assessment framework of rockfall in karst terrains: A case study in Guilin, China

Addressing the insufficient research on rockfall hazard risk assessment in karst regions, this study focuses on the city of Guilin, a typical karst area, to conduct a comprehensive risk evaluation. Rockfalls occur frequently in this region due to complex topography, geological conditions, and intense human activities. The study began by intelligently delineating 60,163 slope units using the circular variance method. A stacking ensemble learning model was employed to integrate the complementary strengths of Random Forest, Extreme Gradient Boosting, Support Vector Machine, and Artificial Neural Network, achieving high-precision prediction of rockfall susceptibility. Subsequently, a comprehensive "hazard-vulnerability" assessment was constructed by incorporating natural and anthropogenic factors via the Analytic Hierarchy Process. Results demonstrated that the stacking ensemble model significantly outperformed individual models and other ensemble methods (blending and weighted averaging). Interpretability of the model was enhanced using SHAP value analysis, which revealed the contributions of driving factors. Uncertainty analysis confirmed the model’s robustness to spatiotemporal variations in the training data. Monte Carlo simulations revealed that the hazard assessment results were stable and reliable. A multi-faceted analysis enhances the credibility and practical value of the final risk map in decision-making. The proposed stacking ensemble framework is methodologically generalizable and provides a robust, adaptable roadmap for rockfall risk assessment in other karst terrains.