Slope stability prediction using automated ensemble learning-based models: a comprensive geotechnical study

Slope stability is a critical concern in geotechnical engineering, yet effective solutions remain elusive due to the complex interactions among various influencing parameters. This study introduces an ensemble learning-based methodology for predicting slope stability, utilising six ensemble learning models within the AutoGluon framework and focusing on key factors such as soil properties and slope geometry. Natural slopes are considered based on a careful investigation of slopes-sources, types of failure and accuracy of field data. The models were trained and validated using a dataset of 157 slope case histories from diverse global regions. Performance was evaluated through metrics including accuracy, precision, recall, the receiver operating characteristic (ROC) curve, and the Area Under the ROC Curve (AUC). Additionally, Shapley Additive Explanations (SHAP) were employed to clarify the contributions of different factors to models predictions. The results indicate that the ensemble learning models performed robustly in predicting slope stability, achieving high accuracy and AUC scores, with the Extra Trees (ET) model identified as the most reliable. SHAP analysis highlighted the influence of contributing factors, closely tied to the selection and preprocessing of input data. Distinguishing between slope types improved both the model’s accuracy and its interpretability. Notably, the friction angle and soil unit weight emerged as primary factors, while water pressure complicated the nonlinear response. Cohesion was significant beyond a threshold. This study establishes a comprehensive geotechnical analysis through careful data selection, removal of misleading information, and identification of significant geological cases, facilitating a thorough examination of the relationship between input data and expected modelling results.