Earthquake-induced landslide susceptibility mapping: Application and Comparison of Frequency Ratio, Logistic Regression, Weight of Evidence and Support Vector Machine

Abstract

This study seeks to use the following GIS-based (Geographic Information System) models for earthquake-triggered landslides susceptibility mapping: frequency ratio (FR), logistic regression (LR), weight of evidence (WE) and support vector machine (SVM). The case of April 20th, 2013 Lushan earthquake in China was employed and analysed by using these GIS-based models to produce landslide susceptibility maps. An inventory containing 1289 landslides related to this earthquake was randomly divided into two parts: 70% for training and the rest 30% for testing. Seven thematic layers representing landslide predictive factors were used in the models to produce the landslide susceptibility maps. The results are validated using two indexes: the relative landslide density (RLI) and area under the receiving operating characteristic (ROC) curves (AUC). Results of RIL showed that in general, all six models were reasonably accurate and the SVM models relatively performed better in earthquake-triggered landslide susceptibility mapping. The AUC values showed that the SVM (RBF) model provided the highest success rate (91.2%) and the second highest prediction rate (79.4%), while the SVM (polynomial) generated the second highest success rate (87.6%) and the highest prediction rate (82.0%). Results using FR model gave the lowest AUC values of 0.727 in success rate and 0.734 in prediction rate. The LR and WE model produced approximate values in success and prediction rate. The resultant landslide susceptibility maps would be helpful for regional planning and reconstruction in this earthquake-prone area.