Landslide Susceptibility Mapping with Support Vector Machine Algorithm

This paper introduces one current machine learning approach for solving spatial modeling problems in domain of landslide susceptibility assessment. The case study addresses NW slopes of Fruška Gora Mountain in Serbia, where landslide activity has been quite substantial, but not inspected in detail. Regarding this lack of precise landslide inventory, an expert-driven zoning of landslide susceptibility was created as the referent model. The Support Vector Machines (SVM) as the last generation machine learning classifier was used to device the susceptibility model after criteria used in the expert-driven model. Training and testing of the SVM algorithm was performed over an assembly of spatial attributes, which included elevation, slope angle, aspect, distance from flows, vegetation, lithology, and rainfall with respect of the referent model. The algorithm was optimized for its learning capacity and kernel dimension parameters. In addition, the autocorrelation precaution was undertaken by specifying the sampling strategy. The results show that high accuracies (87,1%) and agreement coefficients (0,77) between the referent model and the SVM model were feasible with small sample sizes (10% of original instances).