Spatial prediction of groundwater nitrate through machine learning and stakeholder co-creation: the chalk aquifer case in East Anglia, UK

Nitrate contamination in groundwater threatens drinking water supplies worldwide. Data-driven methods can improve prediction and spatial mapping by transforming point observations into continuous regional information, supporting a better understanding of contamination patterns and mitigation strategies. This study presents the co-creation with stakeholders of a continuous nitrate contamination map for the Chalk aquifers of East Anglia, England. Among the classification models, Gradient Boosting and Random Forest performed best, each achieving an accuracy of 0.89. For regression, the Extra Trees model showed the highest performance (R² = 0.64; MAE = 11.0 mg/L). Both regression- and classification-based maps captured nitrate distribution effectively, identifying the unconfined Chalk outcrops in the west and parts of the central and northern valleys as the areas of greatest concern, while confined Chalk aquifers showed lower contamination potential. Both approaches also highlighted unsampled locations outside official nitrate vulnerable zones with characteristics similar to known contaminated areas. Shapley analyses identified geological variables as the strongest predictors of groundwater nitrate. Although both mapping approaches produced similar spatial patterns, they provide complementary information that can support stakeholder decision-making, with improved model interpretability enhancing their practical uptake.