Soft computing in assessment of earthquake-triggered landslide susceptibility

Abstract

This study compared two soft computing methods that are applied to soft computation in assessment of earthquake-triggered landslide susceptibility: the support vector machine (SVM) and artificial neural networks (ANN). As a case study, a series of landslide susceptibility maps were constructed for the affected area of the 2013 Minxian-Zhangxian, Gansu Province, China Mw 5.9 earthquake. From data available, 2330 coseismic landslides were partitioned into two subsets which were used as a training dataset and a test dataset, respectively. In addition, other 1631 points on the map were randomly selected as non-landslide samples in those areas were not influenced by the coseismic landslides. Ten conditioning factors of landslides were considered, including elevation, relative relief, slope angle, slope aspect, slope curvature, slope position, lithology, peak ground acceleration (PGA), distance from the probable seismogenic fault, and distance along the fault. Using the ANN and SVM, ten landslide susceptibility maps were produced for the study area. Cross comparisons and validations of these ten resulting maps with real coseismic landslides show that the polynomial kernel type with a high enough polynomial degree term value (e.g., 5 or 6) of the SVM technology is most appropriate for coseismic landslide susceptibility assessment, which is a challenge to the previously held notion that the SVM method with radial basis function is the most suitable.