Machine learning-supported determination for site-specific natural background values of soil heavy metals.

Heavy metal natural background values play a crucial role in distinguishing anthropogenic sources from natural sources to assess human impacts in polluted areas, thereby accurately formulating environmental policies. However, due to limitations imposed by human activities, research methods, and regional constraints, the determination of heavy metal background values, particularly on site or profile scale, is often challenging, highlighting the urgent need for new methodologies. To establish a comprehensive dataset containing heavy metal concentrations and soil properties, the study systematically collected and screened 82 soil profiles from areas minimally affected by human activities, resulting in a total of 2185 data sets. Using soil depth, pH, organic matter, weathering indices (SAF, BA), FeO, MgO, NaO, CaO, and KO as model input variables, the predictive performance for site-specific background levels of Cd, Cr, Cu, Ni, Pb, and Zn was compared across four advanced machine learning models (RF (random forest), XGBoost (extreme gradient boosting), ANN (artificial neural network), SVR (support vector regression)). The results indicated that the optimal model for predicting background values of Cd, Cr, and Ni was XGBoost (MAE = 0.14 - 0.17; MSE = 0.04 - 0.06; R² = 0.82 - 0.87), while RF was used for Cu, Pb, and Zn (MAE = 0.01 - 0.18; MSE = 0.02 - 0.06; R² = 0.89 - 0.95). Importance assessments using RF and SHAP revealed that pH is a key controlling factor for Cd and Ni, FeO significantly impacts Cr, Cu, and Zn background levels, and KO is the main controlling factor for Pb. The machine learning models developed can effectively predict the background levels of these six heavy metals based on major elemental and soil physicochemical properties, particularly achieving accurate predictions for Cu and Zn using just two input variables. This machine learning prediction framework is based on major elemental compositions and the physical/chemical properties of soil, enables precise and cost-effective point-to-point environmental assessments, thereby offering significant potential for practical applications.