Assessment of GIS-Based Machine Learning Algorithms for Spatial Modeling of Landslide Susceptibility: Case Study in Iran

Abstract

Landslides are one of the most important natural hazards, causing serious financial damages and loss of life in many regions, including Iran. Landslide spatial modeling (LSM) in the current research consists of four phases including: (1) determining the relationship between each conditioning factor and landslide occurrences, using the evidential belief function (EBF) model; (2) utilizing LSM using three geographic information systems (GIS)-based machine-learning models, including the support vector machine (SVM), random forest (RF), and Naïve Bayes (NB); (3) evaluating considered models and selecting the best model using receiver operating characteristics and calculation of the area under the curve (AUC); and (4) feature selection and the rank of importance of conditioning factors using the learning vector quantization (LVQ) algorithm. In order to achieve this aim, first, a landslide inventory map with 146 disaster locations was prepared using national reports and field monitoring, and landslide locations were divided into a training (70%=102) and validating (30%=44) data set. Then, 12 landslide conditioning factors, such as aspect, altitude, drainage density, distance from faults, lithology, slope, land use, plan curvature, profile curvature, distance from rivers, distance from roads, and the topographic wetness index, were selected as input layers for the LSM in the Chahardangeh Watershed. In the next step, the EBF model was used to consider the relationship between landslide locations and the aforementioned conditioning factors. Subsequently, three GIS-based machine learning algorithms were applied for landslide susceptibility mapping, and their results were validated using AUC values. Finally, the LVQ algorithm was applied for feature selection and importance determination of different conditioning factors. The results showed that the RF model had the highest AUC value (83.3%), followed by the SVM, and the NB with AUC values of 75.7% and 72.5%, respectively. Also, the results of the LVQ model revealed that altitude, the distance from the road, lithology, and land use have the most effects on landslide occurrence, respectively. So, the landslide susceptibility maps can be used for land use planning, and the management of landslide hazards in this study area.