Prediction of safety factors for slope stability: comparison of machine learning techniques

Because of the disasters associated with slope failure, the analysis and forecasting of slope stability for geotechnical engineers are crucial. In this work, in order to forecast the factor of safety (FOS) of the slopes, six machine learning techniques of Gaussian process regression (GPR), support vector regression, decision trees, long-short term memory, deep neural networks, and K-nearest neighbors were performed. A total of 327 slope cases in Iran with various geometric and shear strength parameters analyzed by PLAXIS software to evaluate their FOS were employed in the models. The K-fold (K = 5) cross-validation (CV) method was applied to evaluate the performance of models' prediction. Finally, all the models produced acceptable results and almost close to each other. However, the GPR model with R-2 = 0.8139, RMSE = 0.160893, and MAPE = 7.209772% was the most accurate model to predict slope stability. Also, the backward selection method was applied to evaluate the contribution of each parameter in the prediction problem. The results showed that all the features considered in this study have significant contributions to slope stability. However, features phi (friction angle) and gamma (unit weight) were the most effective and least effective parameters on slope stability, respectively.

Automated summary

This study utilized six machine learning techniques to analyze and evaluate 327 slope cases in Iran, with the GPR model identified as the most accurate predictor of slope stability. All features considered in the study made significant contributions to slope stability, with phi (friction angle) and gamma (unit weight) being the most important and least important parameters, respectively.