Integrating Machine Learning for Enhanced Spatial Prediction and Risk Assessment of Soil Heavy Metal(loid)s

Abstract

Accurately predicting the concentrations and spatial distribution of soil heavy metal(loid)s is crucial for effective environmental management and human health risk assessment. However, existing studies are often limited by poor model accuracy, feature selection, and interpretability—particularly under high-dimensional heterogeneous conditions that are inappropriate for more generalised traditional methods. This study proposes an integrated predictive framework combining unsupervised and LASSO-based variable selection, a Lasso-Stacking ensemble model, and SHAP-based interpretability analysis. Using 6,403 soil samples and 34 environmental variables from the arid region of northern China, high-resolution spatial predictions were conducted for 10 heavy metal(loid)s—As, Cd, Co, Cr, Cu, Mn, Ni, Pb, Se, and Zn—alongside ecological and human health risk assessments. The ensemble model significantly outperformed conventional machine learning models, achieving improved prediction accuracy (R² > 0.6) and generalisability. Key environmental drivers influencing the distribution of heavy metal(loid)s included the aridity index, relative humidity, total phosphorus, and bulk density. Spatial analysis revealed that the southern Guanzhong Plain (in Shaanxi) and southern Gansu are hotspots for heavy metal(loid)s, likely affected by both natural and anthropogenic factors. The ecological risk assessment indicated widespread mild contamination by Cd, Se, Pb, and Cu. The health risk analysis revealed high non-carcinogenic risks associated with As, Cr, and Mn in children, and As, Cr, and Ni in both children and adults. This study provides an empirically sound framework for assessing soil pollution risks and supporting targeted environmental management strategies in northern China.