Peak Breach Discharge Prediction for Landslide Dams Using a Bayesian-Optimized XGBoost Model and Sensitivity Analysis

Accurate and rapid assessment of peak flood discharge from landslide dam breaches is crucial for emergency response efforts. Predicting breach parameters of sudden landslide dam failures using machine learning methods has become a current research focus. However, existing landslide dam database slack sufficient case records, and current predictive models for peak breach discharge fail to capture the nonlinear interactions among influencing factors, resulting in limited generalization capability. In response to this, this study employs a sediment erosion model to simulate the landslide dam breach process, thereby expanding the landslide dam breach case database. The Extreme Gradient Boosting (XGBoost) machine learning algorithm is used to predict the peak flood discharge of landslide dam breaches, and the Bayesian algorithm is employed to optimize the hyperparameters of the XGBoost model. An innovative machine learning prediction model for peak flood discharge of heterogeneous landslide dam breaches is proposed, considering eight influencing factors, including geometric parameters of the dam (height, width, length, volume), reservoir capacity, triggering factors, and material composition (erodibility and structural type). The results indicate that, compared to traditional models, the Bayesian-optimized XGBoost machine learning model exhibits higher prediction accuracy. Case analyses of Tangjiashan and Baige landslide dams confirm that the model's predicted peak flood discharge has a maximum error of approximately 20% compared to the actual values. This study provides a valuable reference for emergency response and regional disaster mitigation in the context of landslide dam breaches.