A comparative study of composite kernels for landslide susceptibility mapping: A case study in Yongxin County, China

Abstract

In this study, an effective kernel-based learning framework for landslide susceptibility mapping (LSM) is presented through an implementation of support vector machines (SVMs) with different composite kernels. Kernel-based classification methods are very popular in statistical classification and regression analysis because they can effectively address intractable issues such as the curse of dimensionality, limited known samples and noise corruption. The most representative of such methods is the SVM technique. Although SVMs have recently been widely used in LSM, they were defined using only the attribute value of each influencing factor and did not consider the high dependency between the adjacent vector-valued grid cells. This caused a labelling uncertainty. To solve this problem, it is necessary to combine both the influencing factor's attribute features and spatial dependency information in the SVM. In this work, we present two forms of composite kernels to combine the two aforementioned types of information: 1) constructed through a single kernel with stacked vectors; 2) built through summation kernels under different restrictions. The main advantages of the proposed framework are twofold. First, the integration of the two types of information can improve the predictive capability of the SVMs by removing the isolated class noise in the LSM results. Second, other useful information can be extracted from the spatial domain, such as the structural features of grid cells within and outside of landslide areas. The SVM comparisons were based on data from Yongxin County, China, containing 364 past landslide occurrences that were separated randomly into a training set (70%) and a validation set (30%). The geo-environmental setting of the study area was analysed and sixteen influencing factors were selected. The validation of these SVMs was performed using the receiver operating characteristic (ROC) and the area under the ROC curve (AUC). Experimental results demonstrate that all the SVM-based landslide susceptibility maps have similar spatial distributions upon visual inspection. Specifically, the mountainous zones in the north and south of the study area are characterized by high and very high susceptibility values, respectively, whereas the central part of the study area is categorized as the least susceptible zone. Meanwhile, the composite kernel-based learning framework can achieve a better prediction accuracy than the original SVM. From quantitative analysis, the four SVMs with a summation kernel obtain the AUC values above 0.8900, which is 0.0117 higher than that of the original SVM. Furthermore, a weighted scheme in the summation kernel can result in AUC values that are at least 0.0014 higher than a directional scheme.