Prediction and interpretation of pesticide behavior in acidic soil based on XGBoost-SHAP

The migration and transformation behavior of pesticides in acidic soils is a key link in environmental risk assessment, and its complex nonlinear characteristics pose a challenge to traditional research methods. This review systematically summarizes the research progress of the XGBoost-SHAP interpretable machine learning framework in predicting pesticide behavior in acidic soils, aiming to construct a new analytical paradigm that balances high accuracy and interpretability. The XGBoost algorithm, through ensemble learning mechanism and regularization strategy, can effectively capture the nonlinear response under the coupling effect of multiple environmental factors, and enhance the ability to model the behavior of adsorption, degradation, leaching and other processes; Combining the SHAP interpretation framework and game theory to achieve fair allocation and visualization of feature contribution, the model's predicted results are interpretable in terms of environmental chemistry. At the level of feature construction, research is gradually shifting from a single parameter description to a mechanism guided multidimensional variable system, integrating soil physicochemical properties, pesticide molecular structure, and environmental dynamic factors, and strengthening the model's ability to characterize interface processes such as protonation and complexation. The review further explores the application value of this framework in practical scenarios, including supporting pesticide registration risk assessment, optimizing contaminated site remediation strategies, and guiding regional agricultural precision management. Through global and local interpretation mechanisms, it reveals the pathways and threshold behaviors of key environmental factors. The current research still faces problems such as strong dependence on data quality and insufficient cross category generalization ability. In the future, it can develop towards hybrid architectures that integrate physical mechanism models, spatiotemporal dynamic modeling, and multi-scale interpretation systems. The application of XGBoost-SHAP method marks the transformation of environmental prediction research from black box modeling to interpretable intelligence, providing theoretical basis and technical path for building a scientific and transparent environmental decision support system.