Applicability of Support Vector Machines in Landslide Susceptibility Mapping

Abstract

Landslides in Slovakia are followed by great economic loss and threat to human life. Therefore, implementation of landslides susceptibility models is essential in urban planning. The main purpose of this study is to investigate the possible applicability of Support Vector Machines (SVMs) in landslides susceptibility prediction. We have built a classification problem with two classes, landslides and stable areas, and applied SVMs algorithms in the districts Bytca, Kysucke Nove Mesto and Žilina. A spatial database of landslides areas and geologically stable areas from the State Geological Institute of Dionýz Stur were used to fit SVMs models. Four environmental input parameters, land use, lithology, aspect and slope were used to train support vector machines models. During the training phase, the primal objective was to find optimal sets of kernel parameters by grid search. The linear, polynomial and radial basis function kernels were computed. Together 534 models were trained and tested with LIBLINEAR and LIBSVM libraries. Models were evaluated by Accuracy parameter. Then the Receiver Operating Characteristic (ROC) and landslides susceptibility maps were produced for the best model for every kernel. The best predictive performance was gained by radial basis function kernel. This kernel has also the best generalization ability. The results showed that SVMs employed in the presented study gave promising results with more than 0.90 (the area under the ROC curve (AUC) prediction performance.