An explainable machine learning model for accurate estimation of residual phthalate esters in Chinese agricultural soils

Accurate estimation of diffuse chemical pollution in agricultural soils remains a persistent challenge, particularly for phthalate esters (PAEs), where traditional linear mechanistic models often show biases of 1-2 orders of magnitude. An explainable machine learning framework based on XGBoost was developed to predict residual dibutyl phthalate (DBP) and di-(2-ethylhexyl) phthalate (DEHP) concentrations across Chinese agricultural soils. Integrating the first national-scale dataset on residual plastic film as a key source variable with environmental and anthropogenic drivers reduced systematic prediction bias to coefficients of variation (CV) of + 9.12% for DBP and + 6.57% for DEHP. This significantly outperforms conventional approaches, which exhibited underestimation biases with CV values as low as -99.9%. The robust predictions generated the first high-resolution (1 km) map of PAE distribution in China's farmlands, highlighting elevated concentrations in regions such as eastern Inner Mongolia, central Jilin, northern Shanxi, southern Gansu, eastern Yunnan, and western Guizhou. Through SHapley Additive exPlanations (SHAP) and structural equation modeling (SEM), the underlying mechanisms were clarified. Soil residual plastic not only directly releases PAEs but also indirectly enhances their persistence by promoting soil acidification and reducing cation exchange capacity. This study provides an interpretable, high-fidelity tool for forecasting diffuse chemical pollution in agroecosystems and supports targeted mitigation strategies.