Landslide Susceptibility Assessment in Western External Rif Chain using Machine Learning Methods

Abstract

Landslides are a major natural hazard in the mountainous Rif region of Northern Morocco. This study aims to create and compare landslide susceptibility maps in the Western External Rif Chain context using three advanced machine learning models: Random Forest (RF), Extreme Gradient Boosting (XGBoost), and K-Nearest Neighbors (KNN). The landslide database, created by satellite imagery and field research, contains an inventory of 3528 cases of slope movements. A database of 12 conditioning factors was prepared, including elevation, slope, aspect, curvature, lithology, rainfall, topographic wetness index (TWI), stream power index (SPI), distance to streams, distance to faults, distance to roads, and land cover. The database was randomly divided into training and validation sets at a ratio of 70/30. The predictive capabilities of the models were evaluated using overall accuracy (Acc), area under the receiver operating characteristic curve (AUC), kappa index, and F score measures. The results indicated that RF was the most suitable model for this study area, demonstrating the highest predictive capability (AUC= 0.86) compared to the other models. This aligns with previous landslide studies, which found that ensemble methods like RF and XGBoost offer superior accuracy. The most important causal factors of landslides in the study area were identified as slope, rainfall, and elevation, while the influence rate of TWI and SPI was the minimum. By analyzing a larger inventory of landslides on a more extensive scale, this study aims to improve the accuracy and reliability of landslide predictions in a west Mediterranean morphoclimatic context that encompasses a wide variety of lithologies. The resulting maps can serve as a crucial resource for land use planning and disaster management planning.