A Systematic Comparison of Statistical and Machine-Learning Models for Mapping Landslide Susceptibility: Evidence from the 2018 Rainfall-Induced Landslides in Hiroshima

Landslide susceptibility mapping (LSM) is an essential tool for hazard assessment and land-use planning in landslide-prone areas. This study compares three statistical models—Frequency Ratio (FR), Weight of Evidence (WoE), and Logistic Regression (LR)—with six machine-learning algorithms: Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Artificial Neural Network (ANN), k-Nearest Neighbor (KNN), and Decision Tree (DT), for regional landslide susceptibility assessment in Hiroshima Prefecture, Japan. A balanced dataset comprising 1936 landslide and 1936 non-landslide samples was developed from the 2018 rainfall-induced landslide inventory, utilizing seven conditioning factors: slope angle, profile curvature, aspect, elevation, lithology, soil water index, and 24 h cumulative rainfall. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F1-score. Among the statistical models, WoE exhibited the highest performance, while SVM provided the most balanced results among the machine-learning models. Both modeling approaches consistently identified lithology and slope angle as the primary controls on landslide occurrence. Independent validation demonstrated comparable predictive performance for both models; however, spatial validation showed that WoE assigned 96.72% of observed landslides to the High and Very High susceptibility classes, compared to 72.54% for SVM. These findings underscore the importance of integrating conventional classification metrics with spatial validation to enhance the evaluation and interpretation of landslide susceptibility models for regional hazard assessment.