Landslide susceptibility mapping of the Muş-Bingöl region: a comparative analysis and optimization of machine learning models

Landslides are among the most destructive natural hazards in Türkiye, where high susceptibility is linked to active tectonic structures, steep topography, and complex climatic conditions. For instance, the Eastern Anatolian Region is recognized as a high-risk zone regarding seismicity and vast mass movements; therefore, reliable predictive tools for hazard mitigation are needed. Although several studies have already applied Machine Learning (ML) methodology for Landslide Susceptibility Mapping (LSM) problems in Türkiye, no systematic comparative evaluation of different modelling hierarchies has been performed so far for this particular area in a tectonically complex environment.

This study attempts to fill this gap by developing and rigorously comparing three disparate modeling methods: a statistical baseline, Logistic Regression (LR); an ensemble, Random Forest (RF); and a state-of-the-art deep learning method, Convolutional Neural Networks (CNN). The study was conducted using a landslide inventory and twelve landslide conditioning factor layers, including topographic data: DEM, Slope, Curvature, TWI; geological data: Lithology and Distance to Fault; environmental data: NDVI and Land Cover.

The core methodology embraced a systematic optimization of dataset splitting, whereby model performance was compared across different test/train ratios in order to identify the most stable and accurate data partition. Results are presented using key statistical metrics, including Accuracy and the Area Under the Receiver Operating Characteristic Curve (AUC-ROC), for LR, RF, and CNN. The best-performing model and its corresponding optimal test/train ratio were used to generate the final high-resolution LSM map for the Muş-Bingöl area. This forms a scientifically validated tool that can be used for regional land-use planning and risk management.