Landslide susceptibility mapping using CNN-1D and 2D deep learning algorithms: comparison of their performance at Asir Region, KSA

To be proactive in mountain hazard mitigation, landslide disaster assessments are becoming increasingly urgent. In this study, three modeling techniques, namely, support vector machine (SVM), convolutional neural network (CNN-1D), and (CNN-2D), were applied and their outcomes were compared for landslide susceptibility mapping at Asir Region, Saudi Arabia. As a first step, a landslide inventory map was developed from various data sources. A total of 181 landslide points were identified and divided into 70% training and 30% validation datasets. Thirteen landslide indicator factors (LIFs) were used, including elevation, aspect, distance to fault, geology, land use, plan and profile curvature, distance to road, slope length (LS), stream power index (SPI), topographic witness index (TWI), slope angle, and distance to streams. Experimental results of model accuracy using receiver operating characteristics and area under the curve (ROC, AUC), mean absolute error (MAE), and kappa index (K) showed that the CNN-1D and CNN-2D models (ROC = 86% and 89%, respectively) were more accurate than conventional machine learning model (SVM) (ROC = 82%) in predicting landslides spatially. Specifically, the results showed that CNN-1D and CNN-2D were 4.9% and 7.9% better than support vector machine (SVM) in terms of ROC, and that CNN-2D was 3.5% better than CNN-1D. Moreover, other statistical indices showed that CNN-2D produce the highest value of kappa index (0.855) and lowest value of mean absolute error (0.072), whereas SVM provides the lowest value of kappa index (0.562) and highest value of mean absolute error (0.223). Results indicate that the CNN-2D model is the optimal model for landslide susceptibility mapping. The generated hazard maps are a crucial step in landslide prevention and management to identify the future landslides and avoid potentially problematic areas.