Machine learning-based prediction of deep soil metal(loid) contamination in industrial areas: Role of surface...

Predicting the distribution of soil contamination is crucial for targeted remediation efforts and risk prevention, especially considering the high costs associated with in-situ contamination surveys. This study proposes a random forest (RF)-based approach using readily available surface environmental factors to predict deep soil metal(loid) contamination. The RF model was applied to two smelting areas with relatively homogeneous deep soil parent material. The geomean contents of As, Cd, Cu, Pb, and Zn in these areas exceeded 19.4, 1.34, 90.5, 73.0, and 265 mg/kg, respectively. The deep soil metal(loid) contents exhibited negative correlations with surface soil physicochemical properties such as SOM, Fe, and S content, while positive correlations were observed with metal(loid) and clay contents. Anthropogenic factors, including pollution sources, hydraulic sources, land use, and land cover, also significantly influenced deep soil contamination distribution. The RF model predicted the I_geo distribution of As, Cd, Cu, Pb, and Zn in deep soil, achieving R² values ranging from 0.757 to 0.897. Prediction errors were significantly reduced with increasing surface soil metal(loid) content and showed relatively insensitive to prediction depth. This RF-based approach offers a possible solution for low-carbon soil remediation by substantially reducing the extensive drilling need and energy consumption required for deep soil contamination surveys.