Spatial Prediction Models for Seismic Landslides Based on Support Vector Machine and Varied Kernel Functions: A Case Study of the 14 April 2010 Yushu Earthquake in China

Abstract

AbstractIt is an important subject to study spatial prediction models of landslides triggered by earthquakes in terms of statistics learning theory and Geological Information Systems (GIS) technology. Such models allow us to predict the area of landslide occurrence during an earthquake under similar conditions in the future. On April 14, 2010 a catastrophic earthquake with Mw 6.9 struck Yushu County, Qinghai Province, China. After this event, a total of 2036 landslides were interpreted from aerial photographs and satellite images, verified by field investigations, distributed in a rectangle area of 1455.3 km2. The aim of this study is to carry out spatial prediction modelling of landslides in this rectangle area using several types of kernel functions based on GIS and the support vector machine (SVM) model. A spatial database, including landslides and associated controlling factors which may have influence on the occurrence of landslides, was developed and analyzed using GIS technology. The 12 controlling factors, including elevation, slope angles, slope aspect, slope curvatures, slope positions, drainages, lithology, faults, roads, normalized difference vegetation index (NDVI), co‐seismic surface ruptures, and peak ground acceleration (PGA) were selected as the prediction factors. The landslide spatial prediction mapping were produced by using several types of kernel functions such as linear function, polynomial function, radial basis function (RBF), and sigmoid function based on the SVM model. Then landslide hazard index maps, landslide hazard rank maps, and prediction result maps for the Yushu event were created separately. The success rates of the four types of kernel functions are 79.87%, 83.45%, 84.16%, and 64.62%, respectively. The results show that radial basis function (RBF) is the best kernel function among the four kernel functions for spatial prediction mapping of landslides by the Yushu earthquake. The training and discussion of the model were also carried out based on other three different groups of random distribution training samples and one group of regular distribution training samples. This paper provide an example for selecting appropriate types of kernel functions for prediction mapping of seismic landslides using support vector machine modelling. The hazard rank maps and prediction result maps can be useful in efforts of landslide hazard mitigation.