Annual atrazine residue estimation in Chinese agricultural soils by integrated modeling of machine learning and mechanism-based models.

This study presents a comprehensive approach to estimating annual atrazine residues in China's agricultural soils, integrating machine learning algorithms and mechanism-based models. First, machine learning was used to predict essential parameters influencing atrazine's adsorption, degradation, and dispersivity of solute transport. The results demonstrated that soil organic matter was the most important input variable for predicting adsorption and degradation; clay content was the primary variable for predicting dispersivity. The SHapley Additive exPlanations (SHAP) contribution of various soil properties on target variables were also analyzed to reveal whether each input variable has a positive, negative, or complex effect. Subsequently, these parameters inform the construction of a detailed model across 23,692 subregions of China, with a 20 km × 20 km resolution. The model considered regional variations and soil layer heterogeneity, including rainfall, soil depth-specific properties, and parameters for adsorption, degradation, and dispersivity. Utilizing the convection-dispersion equations and the Phydrus, the model simulated atrazine's transport and degradation patterns across diverse soil environments after applying 250 mL of atrazine (40%) per Chinese mu. The outcomes provided a spatially explicit distribution of atrazine residues, specifying that the arid areas have the highest residual risk, followed by the Northeast, Southwest, and Southeast. Atrazine levels may exceed national drinking water standards at 50 cm depth in Inner Mongolia, the Qinghai-Tibet Plateau, and the Jungar Basin. This study's integrative approach may also offer valuable insights and tools for evaluating residues of various pesticides and herbicides in agricultural soils.