Machine learning-based assessment of three-dimensional slope stability using geometric features under heavy rainfall

Assessment of slope stability under rainfall at the regional scale remains a major challenge in disaster prevention. Traditional geographic information system (GIS)-based methods are efficient but oversimplify slope geometry, while precise numerical simulations are too computationally intensive for large-scale applications. This study applies machine learning (ML) to predict rainfall-induced slope stability at the regional scale. Support vector regression (SVR), extreme gradient boosting (XGBoost), a baseline random forest (RF) model, and an optimized version with recursive feature elimination (RFE-RF) were adopted to capture complex nonlinear relationships between slope geometry and the factor of safety (FoS). Geometric features, including maximum and average slope gradients, surface undulation, and the frequency of undulations exceeding 10 m along the longitudinal direction of the slope, were used as input variables. The proposed framework, integrating numerical simulation and ML, was tested on a highway slope in Nagasaki, Japan. The results indicate that the RFE-RF model achieves the highest predictive accuracy and computational efficiency. With a computation time of only 0.68 s per prediction, this model demonstrates strong potential for real-time and large-scale slope stability evaluations. This approach also enables rapid identification of potentially unstable slope regions for early warning and risk management.