Effectiveness of machine learning algorithms in landslide susceptibility mapping: A case study of Trabzon Province, Turkey

Abstract

Determination of vulnerable zones to landslide is of utmost importance for disaster management and hazard mitigation. Therefore, one of the most significant processes in disaster planning is the production of accurate and up-to-date landslide susceptibility maps. The primary goal of this present work is to investigate the effectiveness of different machine learning algorithms considering random forest (RF), AdaBoost (AB), and logistic regression (LR) for generating landslide susceptibility map of Trabzon province, located in the northeast of Turkey. For this purpose, 12 most widely used landslide-conditioning factors (slope, elevation, plan curvature, profile curvature, slope length, topographical position index, topographical ruggedness index, topographical wetness index, stream power index, normalized difference vegetation index, distance to roads, and distance to rivers) were utilized to produce landslide susceptibility maps and the results were evaluated by utilizing receiver operating characteristic, area under curve (AUC), and overall accuracy (OA) according to confusion matrices. The validation results indicated that AUC obtained using RF, AB, LR methods were computed as 0.929, 0.910, and 0.744, respectively. Additionally, the statistical significance of the methods was evaluated using the McNemar’s test and found that RF and AB methods produced similar results but their results significantly differ from that of the LR method. © 2021 Elsevier B.V., All rights reserved.