Landslide susceptibility mapping based on data mining models in Lesser Caucasus and Kura foreland basin (Armenia and Azerbaijan)

Landslides are a major geological hazard causing significant loss of life and infrastructure damage worldwide. Landslide susceptibility mapping is a crucial, though developing, tool for understanding the spatial distribution of landslide hazard. This study addresses the absence of a comprehensive landslide inventory, limited understanding of causative factors and the lack of regional-scale susceptibility maps for the Lesser Caucasus and Kura Basin (LC-KB). A landslide inventory was created for the Lesser Caucasus of Azerbaijan and compiled with other inventories, documenting 3,659 landslide polygons. Sixteen causative factors were analysed, and multicollinearity tests confirmed no significant correlations. Three Machine Learning (ML) models—Logistic Regression (LGR), Support Vector Machine (SVM) and Extreme Gradient Boosting (XGBoost)—were fine-tuned to create landslide susceptibility maps. Slope is consistently the most influential factor across all models. Results suggest stronger influence of seismic factors than climatic ones. XGBoost achieves the highest accuracy (0.81) on the testing data set, followed by SVM (0.80) and LGR (0.73). The first two models show strong validation performance, with AUC values of 0.89 and 0.87, respectively, while LGR shows a lower AUC of 0.78. The results are vital for planning and disaster management, highlighting areas needing urgent mitigation.