Beyond the source baseline: Model-scale conditional re-expression of soil metal signals across environmental contexts

Source-related information is widely used in soil heavy-metal screening, but source-related conditions do not necessarily coincide with observed concentration patterns. Using 687 cropland soil samples from a tropical monsoon region in southern China, we tested whether soil-property and landscape-hydroclimatic predictors added predictive information beyond a Source-only baseline under nested spatial-block cross-validation. Four XGBoost configurations were compared using spatial out-of-fold (OOF) predictions, grouped SHAP attribution, prediction shifts, and mixed-effects summaries. Source-only OOF R² values were 0.172, 0.151, and 0.147 for Cu, Pb, and As, whereas Full-model values were 0.232, 0.217, and 0.314. Soil-property gains were most consistent for Cu, limited for Pb, and context-dependent for As. Landscape-hydroclimatic prediction shifts were summarized by gradient-interaction, window-dependent, and environment-structured patterns. Samples with SourceSignal (the Source-only OOF prediction) in the upper third were 3.1-3.8 times more likely to fall within the observed P90 tail, yet 35-39% of P90 highs occurred outside this group, and the models with the lowest P90 errors recovered only 47.5-62.3% of mean P90 concentrations. Source-related information therefore provides a useful regional baseline, while source-deviant highs should be prioritized for targeted resampling and local investigation.