Optimizing the Sample Selection of Machine Learning Models for Landslide Susceptibility Prediction Using Information Value Models in the Dabie Mountain Area of Anhui, China

Abstract

The evaluation of landslide susceptibility is of great significance in the prevention and management of geological hazards. The accuracy of the landslide susceptibility prediction model based on machine learning is significantly higher than that of traditional expert knowledge and the conventional mathematical statistics model. The correct and reasonable selection of non-landslide samples in the machine learning model greatly improves the prediction accuracy and reliability of the regional landslide susceptibility model. Focusing on the problem of selecting non-landslide samples in the machine learning model for landslide susceptibility evaluation, this paper proposes a landslide susceptibility evaluation method based on the combination of an information model and machine learning in traditional mathematical statistics. First, the influence factors for landslide susceptibility evaluation are screened by the correlation analysis method. Second, the information value model is used to delimit areas with low and relatively low landslide susceptibility, and non-landslide points are randomly selected. Third, a landslide susceptibility evaluation method combined with IV-ML, such as logistic regression (IV-LR), random forest (IV-RF), support vector machine (IV-SVM), and artificial neural network (IV-ANN), is established. Finally, the landslide susceptibility factors in the Dabie Mountain area of Anhui Province are analyzed, and the accuracy of the landslide susceptibility evaluation results using the IV-LR, IV-RF, IV-SVM, and IV-ANN and LR, RF, SVM, and ANN methods are compared. The accuracy is evaluated by examining the ACC, AUC, and kappa values of the model. The results indicate that the evaluation effect of the IV-ML models (IV-LR, IV-RF, IV-SVM, IV-ANN) on landslide susceptibility is significantly higher than that of the ML models (LR, RF, SVM, ANN).