Agricultural legacies drive groundwater nitrate contamination: Machine learning and risk-informed monitoring optimization in the Baiyangdian Lake watershed

Groundwater nitrate contamination poses a significant threat to water security in intensively agricultural regions. Effective management is hampered by an incomplete understanding of the spatiotemporal dynamics, particularly the legacy effects of historical farming and the modulating role of the vadose zone. This study presents a comprehensive analysis in the Baiyangdian Lake Watershed, a typical agricultural region in the North China Plain (NCP), leveraging a unique dataset with 20 years of crop changes, detailed vadose zone stratigraphy, and 278 groundwater samples. An explainable machine learning framework (XGBoost-SHAP) was employed to quantify the relative importance of historical agricultural practices and vadose zone properties, enabling the development of a SHAP-WHO Combined Risk Index (SWRI) that integrates predictions with feature contributions for high-resolution risk mapping. The model revealed a multi-decadal lag, with early-2000s irrigation and fertilization rates being dominant predictors of current nitrate levels. Critical thresholds were identified (> 300 mm/y irrigation; > 265 kg/ha/y N application; > 16 m water table depth) beyond which contamination risk sharply increases. Vadose zone properties (particularly the thickness of Holocene deposits) and long-term average water table depth were critical controls on nitrate transport and attenuation. SWRI demonstrated strong predictive power and was used to optimize the groundwater monitoring network via a risk-weighted multi-objective spatial clustering algorithm. This enabled a 70.5% reduction in monitoring sites while maintaining complete coverage of very high-risk areas. This research provides a mechanistic understanding of legacy nitrate contamination and a generalizable, data-driven framework for sustainable groundwater management globally.