Spatial modeling and susceptibility zonation of landslides using random forest, naïve bayes and K-nearest neighbor in a complicated terrain

• Published in: Earth science informatics 2021-09, Vol.14 (3), p.1227-1243
• Main Authors: Abu El-Magd, Sherif Ahmed, Ali, Sk Ajim, Pham, Quoc Bao
• Format: Article
• Language: English
• Subjects: Algorithms; Anthropogenic factors; Earth and Environmental Science; Earth Sciences; Earth System Sciences; Elevation; Hazards; Information management; Information Systems Applications (incl.Internet); Landslides; Landslides & mudslides; Machine learning; Model accuracy; Ontology; Remote sensing; Research Article; Roads; Simulation and Modeling; Slopes; Space Exploration and Astronautics; Space Sciences (including Extraterrestrial Physics; Tributaries; Wetness index; Zonation
• ISSN: 1865-0473; 1865-0481
• DOI: 10.1007/s12145-021-00653-y

Recently, one of the most frequent natural hazards around several regions in the world is the landslide events. The area of Jabal Farasan in the northwest Jeddah of Saudi Arabia suffers from landslide events. The main cause of these events was identified due to the anthropogenic activities represented by mining activities. In this work, different machine learning algorithms (MLA), Random Forest (RF), K-Nearest Neighbor (KNN), and Naïve Bayes (NB) were implemented to predict the landslides events integrated with remote sensing data. The objective is to generate landslides susceptibility prediction map using different MLA. The landslides inventory map was prepared in the present study based on the historical landslide’s events. Landslides at 1354 locations were used to train the models and validate the prediction. Landslides controlling factors include: elevation, slope, curvature, aspect, distance from roads, lineaments density, and topographic wetness index (TWI). Our findings indicate that landslides more likely occurs in the areas of mining activities, close to the roads and the Wadi tributaries due to the high slope angle in some cases. Subsequently, the prediction maps were classified into landslide occurrence location and non-landslides occurrence’s model's validation using the receiver operating characteristic (ROC) curve showed that the model accuracy varied between 86 and 89% for RF, KNN, and NB. The produced landslide susceptibility map in this study would provide useful information for hazard management and control in such natural hazards.