Landslide susceptibility mapping using machine learning algorithms

In this study, it was aimed to carry out landslide susceptibility mapping of the area using remote sensing and geographical information systems in a 38.54 km2 area within the borders of Kartal District located in the south of the Anatolian side of Istanbul. The study area has an important location due to its population and trade area. Within the scope of the study, topographic, morphometric and spectral parameters created from the different data set were used. These parameters includevisible and near infrared, short wavelength infrared, thermal infrared, normalized difference vegetation index, decorrelation stress, principal component analysis, digital elevation model and its derivatives (slope, aspect, distance to canals, plan curvature, profile curvature, topographic wetness index, LS factor), distance to existing fault lines and landslide inventory data obtained from ASTER satellite images and their analysis. In the study, landslide susceptibility mapping was performed by using the input parameters obtained from the analysis and processing results and the Support Vector Machine (SVM), one of the machine learning algorithms. In the study, training and test classes were determined using the Two Level Random Sampling (2LRS) algorithm. The most important feature of this algorithm is the use of pre-landslide conditions in the study area. Thus, landslide susceptibility mapping was performed, not automatic landslide mapping. In the landslide susceptibility map produced, each pixel is classified by showing continuous susceptibility values between 0.0 and 1.0. The resulting susceptibility maps are categorized into four different classes, each of which is high, medium, low and very low susceptibility. Receiver operating characteristic (ROC) curve and area under the curve (AUC) values were used to evaluate the accuracy of the maps obtained as a result of classification. AUC values calculated as the accuracy values of the produced landslide susceptibility maps were found as 93%.