Integrating Multi-Source Environmental Data to Identify Drivers and Assess Risks of Heavy Metal Contamination in Agricultural Soils

The pronounced spatial heterogeneity of soil heavy metals, coupled with the high costs and intensive resource requirements associated with large-scale sampling, presents substantial challenges for accurate spatial prediction and risk assessment. Although conventional machine learning methods demonstrate strong predictive capabilities, they often struggle to simultaneously address spatial non-stationarity, complex nonlinear relationships, and limited model interpretability. By integrating multi-source datasets with multiple machine learning models, SHAP analysis, Monte Carlo simulations, and bivariate Moran’s I analysis, this study developed an integrated and interpretable framework for spatial prediction and risk assessment. The results revealed that (1) the proposed framework achieved high predictive accuracy, with coefficients of determination (R2) exceeding 0.80 for arsenic (As), cadmium (Cd), copper (Cu), lead (Pb), chromium (Cr), and zinc (Zn). Climatic conditions and soil properties were identified as the dominant factors controlling the spatial distributions of heavy metals; (2) As and Cr were identified as the major contributors to carcinogenic risks, whereas children exhibited higher overall health risks than adults throughout the study area; (3) spatial hotspot analysis revealed significant heavy metal pollution hotspots in Gongzhuling City, northern Yitong County, and Dongliao County. This study provides new insights into soil heavy metal exposure and risk assessment, offering scientific support for targeted environmental management and pollution prevention strategies.