Mapping soil salinity and prediction uncertainty across different depths using quantile regression forests

Accurate assessment of soil salinity across the soil profile is necessary for sustainable agricultural management, yet conventional surface-based observations often provide limited information about subsurface salinity conditions. This study developed a depth-resolved digital soil mapping framework to predict soil electrical conductivity (EC) across five standard depth intervals (0–25, 25–50, 50–75, 75–100, and 100–125 cm) in Shandong Province, eastern China. A total of 617 soil profiles were used, and continuous profile observations were harmonized into standard layers using equal-area quadratic spline functions. Environmental covariates representing spectral salinity characteristics, vegetation conditions, climate, and terrain attributes were compiled at a 30-m spatial resolution. Principal component analysis was applied to reduce multicollinearity among predictors, and Quantile Regression Forest (QRF) models were implemented to estimate EC and its 90% prediction intervals. Model accuracy was evaluated using repeated 10-fold cross-validation. The results showed a clear decline in predictive performance with increasing depth: the coefficient of determination decreased from 0.79 in the 0–25 cm layer to 0.51 in the 100–125 cm layer, whereas RMSE increased from 0.142 to 0.295 dS m⁻¹. Surface EC predictions were mainly driven by salinity-sensitive spectral indices, including SSIB, CSI, and SC, while the contribution of these variables weakened in deeper layers. Conversely, topographic wetness index, valley depth, relative slope position, and climatic variables became increasingly important below 50 cm, suggesting a shift from predominantly spectral controls near the surface to hydro-topographic and climatic regulation at depth. Higher EC values were generally predicted in the eastern and southeastern parts of the study area, whereas uncertainty was more pronounced in western and central transitional zones. The findings demonstrate the value of integrating depth-specific environmental information with probabilistic machine learning to produce spatially explicit, uncertainty-aware maps of soil salinity for improved monitoring and site-specific land management.