Sampling effects on machine-learning performance and tectonic controls on landslide susceptibility: insights from the Adra River basin (SE Spain)

This study develops an improved methodological framework for landslide susceptibility assessment in the Adra River basin (SE Spain), where landslides affect nearly 13% of the territory and their spatial distribution is strongly controlled by geology and tectonics. While single-point sampling is effective when inventories are dominated by small-sized landslides, the Adra inventory includes highly heterogeneous failures—from large DGSDs to small shallow slides—requiring adapted sampling. We compared centroid-based sampling with proportional stratified sampling (PSS), combined with balanced (1:1) and background-enriched (1:10) ratios. Nineteen landslide conditioning factors (LCFs), derived from topographic, geological, climatic and land-use data, were reduced to 14 through feature selection. Seven machine-learning models (ANN, GAM, GWLR, MaxEnt, MARS, RF, and SVM) and a mean-ensemble configuration were tested. The GWLR-Mean-PSS ensemble performs best, particularly by balancing 1:1 and 1:10 ratios to achieve high predictive accuracy (AUC = 0.960, 95% CI: 0.954–0.966; Accuracy = 89%) while mitigating spatial bias and ensuring Degree of Fit (DF) of 96.69%. These susceptibility patterns are consistent with the basin’s geodynamics, showing a strong relationship between inherited structural controls, ongoing Quaternary uplift and the landslide density. This validates the model's ability to integrate local tectonic signals with stable predictive power in complex mountainous terrains.