Interpretable soil thickness mapping from sparse profiles through convergent feature selection in the Angulinao semi-arid lake basin

Soil thickness is highly heterogeneous in arid and semi-arid landscapes, while profile-based surveys are costly and often provide only limited observations. Predicting soil thickness from sparse profiles may also be prone to overfitting when numerous environmental covariates are available. We evaluated an interpretable soil-thickness mapping workflow in the Angulinao semi-arid lake-basin landscape using 92 observations and 27 covariates representing terrain, hydrology, climate, remote sensing, and soil-forming background. Boruta and recursive feature elimination converged on the same seven-variable subset: Relief15, TPI21, TPI3, TWI, Slope, mean annual precipitation (MAP), and Elevation. Under leave-one-out cross-validation using predefined covariate subsets and fixed model configurations, this subset combined with XGBoost achieved the highest performance among the tested combinations, with R² = 0.759, RMSE = 24.265 cm, and MAE = 17.494 cm. Its residuals showed no significant spatial autocorrelation using a five-nearest-neighbour Moran’s I test (I = 0.037, p = 0.217). SHAP analysis indicated that model output was associated mainly with Elevation, Slope, Relief15, and water-related covariates. These results demonstrate how convergent feature selection, model comparison, residual spatial diagnosis, and model interpretation can support soil-thickness mapping from limited profile observations.