An evaluation of SVM using polygon-based random sampling in landslide susceptibility mapping: The Candir catchment area (western Antalya, Turkey)

The main purpose of this study was to present an approach that uses all of the input parameters from remotely sensed data to map landslide susceptibility. Furthermore, a novel sampling strategy, namely polygon-based random sampling (PBRS), which maintains the complete independence of sampled data sets for training and testing, was proposed to generate more realistic landslide susceptibility maps. An ASTER image of the Candir catchment area which is located in western Antalya (Turkey) was selected for implementing the proposed approach using a support vector machine classification (SVM) algorithm. The proposed methodology contains three sections: a polygon based sampling algorithm, an SVM classification, and an accuracy assessment. Two data sets (A and B) were generated and compared. Topographical parameters, proximity parameters and Normalized Difference Vegetation Index (NDVI) were used in the two data sets. In addition to these common parameters, data set (A) included lithological unit data produced from conventional geology maps and data set (B) had decorrelation stretched ASTER bands with four mineral (alunite, kaolinite, calcite, and quartz) indices. To construct and evaluate the models, training and testing data sets were generated using the proposed sampling strategy with three random sets for each data set (A and B). Next, the spatial performance of the obtained landslide susceptibility maps was evaluated using the area under the receiver-operating characteristic curves (AUC). The AUC values of the three random sets from data set (A) were 0.913, 0.912, and 0.906. The AUC values of the three random sets from data set (B) were 0.923, 0.912, and 0.907. After a comparison of the obtained AUC values, data set (B) presented considerably acceptable spatial performances in landslide susceptibility map production.