Runout Assessment of Rainfall-Induced Landslides by Coupling BO-XGBoost and Continuum–Discontinuum Simulation

Abstract

Reliable prediction of landslide runout distance and influencing areas is critical for landslide hazard zoning and risk assessment. This paper proposed a rainfall-induced landslide prediction and simulation platform based on the developed interpretable Bayesian optimization-extreme gradient Boosting (BO-XGBoost) model and the continuum–discontinuum element method (CDEM) incorporated with a particle strength softening model. First, based on kinematic prediction theory, several key influencing features were selected from rainfall-induced landslide cases to train the BO-XGBoost model. The nonlinear relationship between runout distance and influencing features was effectively captured by the model, with an R2 of 0.858. According to Shapley Additive Explanations (SHAP) analysis, the elevation difference between the rear and front edges was identified as the dominant controlling feature. Then, the particle flow method was introduced into CDEM and a particle strength softening model was proposed to simulate the landslide process. The applicability of the proposed softening model was validated through the case of the Yigong landslide. Additionally, a collaborative application of the BO-XGBoost model and CDEM simulation was implemented for the Wangjiawan landslide. The runout distances predicted by the BO-XGBoost model and CDEM simulation were 256.8 m and 320 m, respectively, corresponding to relative errors of 14.40% and 6.67% compared with the observed value of 300 m. The methodological innovation and application results demonstrate that this paper provides a reliable reference for the assessment and mechanism analysis of landslides.