Prediction of Slope Stability Using Four Supervised Learning Methods

Abstract

The analysis and prediction of slope stability are very important, because slope failure can lead to large disasters. This paper focused on a performance comparison of four supervised learning methods for slope stability prediction. Based on characteristics of slope instability and analysis of data availability, six typical slope parameters—the unit weight, cohesion, internal friction angle, slope inclination, slope height, and pore water ratio—were chosen to establish the evaluation index system. The gravitational search algorithm (GSA), random forest (RF), support vector machine, and naive Bayesian (Bayes) were proposed to establish classifiers. A data set from more than 10 domestic and abroad slope projects was established to train and test the four classifiers, and then, key parameters of the four models were optimized by using the method of 10-fold cross validation. The prediction performances of the four supervised learning methods were compared and analyzed. The results of accuracy, Kappa, and receiver operating characteristic curves reveal that both GSA and RF models can achieve satisfactory results, and the GSA model can obtain the best results when compared with the other three learning methods. Finally, seven models with varying indicators are investigated to obtain the parameter sensitivity based on RF and GSA models.