Rainfall instability prediction for three types of landslide-prone slopes using various machine learning algorithms

How to quickly and accurately predict the stability of easily sliding slopes under sudden rainfall events is crucial for safe production and operation. However, most existing machine learning (ML) models have not fully accounted for the influence of rainfall factors. Thereby, a new slope rainfall instability prediction scheme based on Latin Hypercube Sampling (LHS), SEEP/W, SLOPE/W, and ML is proposed. By integrating LHS with SEEP/W and SLOPE/W simulations, a comprehensive database incorporating geometric, mechanical, and rainfall parameters was established for three typical slope types in Guizhou Province, China: homogeneous slopes, accumulation-layer slopes, and coal-measure strata slopes. Based on this database, eight ML algorithms were applied to predict slope stability. The results show that the Extreme Gradient Boosting achieved the best performance (average AUC = 0.975), while the Artificial Neural Network performed the worst (average AUC = 0.910). Furthermore, feature importance analysis was conducted to identify key controlling factors and optimize the model inputs, resulting in an overall AUC improvement of approximately 0.1%–4.3%. The reduced-input models were applied to three real cases, achieving average prediction errors of approximately 25% for the Fuxin slope in Fuxin, Liaoning Province, 12.5%–25% for the H1 landslide along the Qinglong–Xingyi Expressway in Guizhou Province, and 12.5%–37.5% for the Maoshajing slope in Guiyang, Guizhou Province, under different rainfall durations. The proposed rapid prediction method for rainfall-induced slope instability can provide technical support for predicting slope instability disasters in geotechnical engineering.