A comparative study of the bivariate, multivariate and machine-learning-based statistical models for landslide susceptibility mapping in a seismic-prone region in China

Abstract

Statistical landslide susceptibility mapping (LSM) models have been most widely used in literatures. However, limitations and uncertainties remain in these methods. The main goal of the current study was to test and compare the efficiency of a bivariate model (the weight of evidence (WoE)), a multivariate model (logistic regression (LR)) and a machine-learning algorithm (the support vector machine (SVM)) in LSM. Lushan County of China was chosen because of its mountainous terrain and high risky of devastating seismic activities. An inventory of 867 landslides was utilized in this study, 70% of which were used to train these models, and the rest 30% were used to validate their accuracies. Ten factors of aspect, elevation, slope, curvature, peak ground acceleration (PGA), distance to the river (DtoR), lithology, topographic wetness index (TWI), stream power index (SPI) and percentage of tree cover (PTC) were used as input of the landslide susceptibility mapping (LSM) models. Accuracy evaluation based on the areas under the receiver operating characteristic curves (AUC) showed that the LR model gives the highest success rate (78.2%) and prediction rate (76.4%), the SVM has the second-highest success rate (75.9%) and the WoE had the second-highest prediction rate (75.6%). Comparison results suggested that the LR and the SVM are proper models for LSM of the study area. The obtained susceptibility maps would benefit regional land planning and seismic landslide hazard mitigation in the study area.