Abstract
methods
Logistic Regression (LR), Support Vector Machines (SVM), and Deep Learning (DL) to identify areas most susceptible to landslides. The selection of causative factors was based on a detailed statistical study examining the relationship between landslide occurrence and specific characteristics such as slope, lithology, Normalized Difference Vegetation Index (NDVI), Topographic Wetness Index (TWI), land use, proximity to roads, watercourses, and geological faults. These factors were essential in generating accurate and reliable susceptibility maps using Geographic Information Systems (GIS) technology. Metrics of performance, including accuracy, precision, F1-score, specificity, sensitivity, and RMSE, were used to evaluate the performance of the models, which were verified, validated, and compared using the area under curve (AUC) value of the Receiver Operating Characteristics Curves (ROC) method and the spatial validation technique. This spatial validation evaluated the percentage of active landslide areas in the high and very high susceptibility classes. The DL and SVM models demonstrated a very high concentration of landslide points in these classes, with 99% and 98% respectively, whereas the LR model showed 89%. In terms of AUC validation, the DL model achieved the highest AUC value of 0.9894, followed by the SVM model followed with an AUC of 0.9873, while LR demonstrated a lower AUC of 0.9093. These precise and reliable results help to identify high-risk areas more effectively, thereby safeguarding residents and preserving infrastructure in the Oued Guebli watershed. The choice of the DL model as the most effective method underscores its capability to deliver accurate susceptibility maps, which are important for informed decision-making and risk management.