A hybrid data-driven machine learning method for mapping trace metals in urban soils: Integrating source...

Mapping regional soil pollution with limited sampling points is essential for effective soil environmental management. The high heterogeneity of urban environments leads to considerable spatial variability in soil trace metals, posing challenges for accurate prediction using conventional methods. This study proposes a Hybrid Data-Driven Machine Learning Method (HDML) for high-precision soil trace metal distribution mapping, which integrates environmental factors, spatial distribution trends, and source apportionment information. In the case study of the Chang-Zhu-Tan urban agglomeration in China, the HDML demonstrated high predictive performance for Cd, As, Pb, and Cr, with R² values ranging from 0.796 to 0.883, though lower for Mn (R² = 0.669). The mean RMSE of the HDML-XGB model was reduced by 77.89% and 51.55% compared to Kriging and XGB, respectively. Spatial distribution trends and source apportionment information contributed over 68% of the total feature importance, exhibiting higher explanatory power for trace metal distribution than environmental factors, such as soil physicochemical properties, climate conditions, air quality, and anthropogenic factors. The HDML effectively minimized prediction errors at both low- and high-value sites and captured more spatial distribution details. By leveraging the multi-source information, the HDML offers a reliable tool for targeted remediation and risk management.