Interpretable soil thickness mapping with nested and spatial validation in the Angulinao semi-arid lake basin

Soil thickness varies markedly across semi-arid basins, yet sparse profile observations make these spatial patterns and their environmental controls difficult to resolve. We mapped soil thickness in the Angulinao semi-arid lake basin using 92 profiles and 27 environmental covariates. Twelve feature-selection and regression pipelines were evaluated by repeated nested cross-validation. Boruta-random forest achieved the lowest mean root mean square error (RMSE; 25.941 ± 0.859 cm), while elevation, slope, relief, topographic position, wetness and precipitation recurred most consistently across training subsets. A fixed seven-variable XGBoost model achieved the lowest RMSE in the five-zone internal spatial comparison (26.358 cm; R² = 0.716) and was used for regional mapping. Thicker soils were concentrated across broad basin lowlands, whereas dissected terrain showed greater spatial heterogeneity; the > 100–150 cm class covered 48.30% of the 3318.773 km² mapped domain. SHAP analysis identified elevation as the dominant model contribution and revealed conditional effects across topographic scales. Bootstrap refits produced a mean pixel-wise standard deviation of 10.072 cm, and 3.16% of the mapped area fell outside the model-defined area of applicability. The resulting maps characterize basin-scale soil-thickness patterns, their environmental associations and priority areas for field verification, providing spatial information for regional soil-resource assessment.