Machine Learning Techniques to Predict Slope Failures in Uttarkashi, Uttarakhand (India)

Abstract

Uttarkashi region is highly prone to landslides because of its geological structure. The exact occurrence of these landslides events is difficult to predict due to its complex mechanism and dependence on the number of triggering factors. Moreover, the behavior and prediction of unstable slopes are of high importance failing of which otherwise can have a devastating impact. This research work aims at studying and modeling slopes with the help of state-of-the-art machine learning techniques: Support vector machine, Backpropagation, Decision tree, and Bayesian Network models. To train and test these models a total of 137 instances are taken. Moreover, the independence of individual features is studied with the help of multicollinearity analysis. The capability of considered methods was evaluated using various performance evaluation metrics. Evaluation and comparison of the results show that the performance of all classifiers is satisfactory for slope failure analysis (AUC=0.750-0.889). Based on the results Bayesian network proved to be most efficient to predict slope failures (Accuracy=93%, AUC=.889). These outcomes can be beneficial for government agencies in early-stage risk mitigation.