Fine-scale mapping and nonlinear attribution of heavy metal(loid)s in urban soils using spatially enhanced machine learning algorithms

Sparse sampling and spatial heterogeneity constrain the prediction of heavy metal(loid)s in urban soils. Using 77 composite surface-soil samples from Hangzhou, China, we compared Random Forest, XGBoost, and Support Vector Regression (SVR) for eight elements under three predictor schemes: environmental variables (ENV), ENV with local spatial lags (ENV + SL), and ENV + SL with interpolation-derived predictors (ENV + SL + SP). Standalone interpolation and ENV-based regression kriging provided additional comparisons. Selected models achieved pooled out-of-fold R 2 values of 0.322–0.555, with the lowest value for Cd and the highest for Pb. XGBoost performed best for six elements, whereas SVR performed best for Mn and Pb. ENV + SL + SP was selected for seven elements, while Pb was best predicted by ENV-only SVR. The best ENV + SL + SP models exceeded the stronger ordinary kriging or inverse distance weighting baseline by 0.040–0.343 in R 2 . Residual kriging improved ENV-only predictions in 15 of 24 comparisons. ENV-only SHAP analysis identified element-specific associations with soil properties, vegetation, precipitation, and road proximity. Predictions were generated for 82,782 grid cells at 250 m resolution. Conditional residual-bootstrap ranges for background-exceedance area were approximately 79–85% for As, 83–91% for Cd, 52–62% for Cu, and 43–54% for Zn. Based on out-of-fold exceedance detection, maps for As, Cu, and Zn may help prioritize confirmatory sampling, whereas those for Cd, Cr, Mn, Ni, and Pb should primarily support preliminary screening. Independent field validation remains necessary, and low predicted concentrations cannot exclude localized hotspots.