Comparative Machine Learning Approaches to Landslide Susceptibility in Southeastern Türkiye

Abstract

Landslides are among the most destructive natural hazards in mountainous regions, causing severe economic losses and infrastructure damage. Reliable susceptibility mapping is essential for hazard mitigation and sustainable land-use planning, particularly in tectonically active areas such as southeastern Türkiye. This study aims to compare the predictive performance and spatial consistency of six widely used machine learning algorithms—Support Vector Machine (SVM), Logistic Regression (LR), Random Forest (RF), Artificial Neural Network (ANN), Naïve Bayes (NB), and Maximum Entropy (MaxEnt)—together with an Ensemble model, under identical geomorphological and environmental conditions. A GIS-based framework was applied using eleven conditioning factors derived from topographic, geological, hydrological, and land-cover datasets. A landslide inventory verified by field observations was used for model training and validation. Model performance was assessed using Receiver Operating Characteristic (ROC) analysis and the Area Under the Curve (AUC). All models achieved high predictive accuracy (AUC = 0.877–0.959). The NB model performed best (AUC = 0.959), while the Ensemble model produced the most spatially consistent susceptibility patterns. High- and very high-susceptibility zones were consistently associated with steep slopes and weak lithological units. Comparative machine learning approaches provide reliable and transferable tools for regional landslide susceptibility assessment. The results highlight the importance of algorithm selection and ensemble strategies in susceptibility mapping.