Interpretable Machine Learning Model for Karst Collapse Susceptibility in Tunnel-Dense Mountainous Areas: A Case Study of Zhongliang Mountain

• 1、School of Civil Engineering, Chongqing University

Abstract：This study aims to develop an interpretable machine learning model for karst collapse susceptibility assessment in tunnel-dense mountainous areas. Taking Zhongliang Mountain as a case study, a dataset comprising 424 positive samples (collapse locations) and 2544 negative samples was constructed. A total of 20 core conditioning factors were selected through the dual application of the Geodetector and Recursive Feature Elimination (RFE) methods. Subsequently, original and Bayesian-optimized Random Forest (RF) and XGBoost models were developed. The evaluation results demonstrate that the Bayesian-optimized XGBoost (BO-XGBoost) model achieved the optimal comprehensive performance, yielding the highest accuracy (0.9877), recall (0.9688), and Kappa coefficient (0.9503), alongside an outstanding Area Under the Curve (AUC) of 0.9983, marginally second only to the unoptimized XGBoost (AUC = 0.9985). Based on the BO-XGBoost model, a five-tier susceptibility zonation map was generated. Furthermore, the SHapley Additive exPlanations (SHAP) approach was introduced to reveal the core conditioning factors across different susceptibility levels. The findings indicate that karst collapses in the study area are the coupled outcome of topographical conditions and tunneling disturbances. The extremely high-risk zones are prominently distributed along the tunnel corridors. This research provides a valuable methodological reference for targeted disaster prevention and mitigation along tunnel alignments in mountainous regions.

Keywords： Karst collapse Susceptibility assessment XGBoost Random Forest Bayesian optimization SHAP interpretability Zhongliang Mountain