From Data Design to Map Footprint: Quantifying Sampling Effects on Landslide Susceptibility

This study evaluates how non-landslide (background) sampling strategy and the landslide: non-landslide (LS: NLS) ratio influence the performance and reliability of landslide susceptibility models.Five learners-Logistic Regression (LR), Random Forest (RF), Extreme Gradient Boosting (XGB), Multilayer Perceptron (MLP), and a Convolutional Neural Network (CNN)-were trained and tested under five dataset regimes.These differed by sampling design (random vs. selective/curated), representation density (single point vs. multi-point per landslide polygon), and LS:NLS ratio (1:1 and 1:3).A unified pipeline was implemented in which preprocessing, stratified crossvalidation, isotonic probability calibration, and Youden-J threshold optimization were held constant across models.Model performance was assessed using ROC-AUC, accuracy, precision, recall, κ-coefficient, and reliability metrics including the Brier score.Spatial leakage was quantified through distance-to-nearest training point statistics to evaluate potential over-optimism caused by clustered sampling.Results show that curated background sampling and balanced class ratios (particularly 1:1 datasets) improved calibration and reduced operating-point bias, while random or imbalanced designs required more permissive decision thresholds to maintain recall.Despite differences in learning architecture, the models converged on physically meaningful controls, notably slope-aspect-relief interactions, proximity to roads and faults, soil type, and antecedent moisture conditions.Basin-scale hydrological indices contributed to smaller but systematic gains.Overall, the findings demonstrate that sampling design exerts as much influence on susceptibility mapping as the choice of algorithm, and that calibrated probabilities are essential for producing reliable and operationally useful landslide susceptibility maps.The study highlights the need for transparent reporting of sampling regimes, spatial independence diagnostics, and probability calibration when developing and comparing landslide susceptibility models.