Study on the Performance of Machine Learning-based Models for Slope Stability Prediction

Slope stability is a critical influencing factor for engineering geological problems such as landslides and subgrade collapses. This study is based on 168 sets of measured slope data in the literature, with slope height, slope inclination angle, unit weight, cohesion, internal friction angle, and pore pressure ratio as input parameters, and stable/failure as the prediction objective. Four machine learning models, LightGBM, XGBoost, SVM, and MLP, were used for prediction and analysis. The correlation between input features was evaluated by Spearman correlation analysis. Subsequently, the prediction performance of models was evaluated by F1 score, accuracy, and AUC value. The model performance evaluation revealed that the ensemble learning models (XGBoost and LightGBM) outperformed than SVM and MLP models. Specifically, XGBoost made the slightly higher AUC value (0.997) and F1 score (0.971) on the testing set while compared to LightGBM (AUC=0.985, F1=0.963). Feature importance analysis revealed that slope height and slope angle have the greatest impact on stability, followed by cohesion and internal friction angle. Finally, the partial dependency graph of the XGBoost model was used to further reveal the influence of various input parameters on slope stability.