A comparison of different machine learning models for landslide susceptibility mapping in Rize (Türkiye)

Abstract

The main purpose of this study was to compare the performance and validation of six machine learning models (extreme gradient boosting, random forest, artificial neural network, support vector machine, C4.5 decision tree, and naive Bayes) in landslide susceptibility modelling. The province of Rize, which has the highest rate of landslide events in Türkiye, was chosen as the study area. The conditioning factors (distance to roads, lithology, drainage density, slope, topographic wetness index (TWI), soil depth, distance to rivers, land use, NDVI, plan curvature, elevation, aspect, profile curvature) affecting the landslide were determined using the ReliefF method. A total of 516 landslides were identified for creating models, comparing performance, and validating results. The performance and validation of the models were determined by the receiver operating characteristics (ROC), sensitivity, specificity, accuracy, and kappa index. The results show that the XGBoost model outperforms the other five machine learning models in terms of accuracy and performance and is the most effective model for generating landslide susceptibility maps in Rize (Türkiye).