A machine learning approach to estimating nitrate-nitrogen concentrations in shallow groundwater across New Zealand

Nitrate contamination of shallow groundwater is a concern for drinking water safety and ecosystem health globally, yet national-scale prediction techniques remain limited. Here, we describe the development and evaluation of an ensemble machine learning framework to estimate median nitrate (NO₃-N) concentrations in shallow groundwater across New Zealand. From 957 state of the environment monitoring bores, 589 shallow (<50 m), oxic bores with complete model data were retained. These were combined with 99 predictor variables describing land use, hydrogeology, climate, topography, and soils. Candidate tree-based machine learning models (RF, XRT, XGBoost, LightGBM) were trained and stacked into an ensemble, evaluated with 10-fold cross-validation repeated 10 times within the training set. On the held-out 20% test set, the ensemble model achieved an RMSE of 3.13 mg/L NO₃-N and R² of 0.56, indicating moderate performance. Soil age, dairy cattle density, land use, and bore depth were amongst the strongest predictors of median groundwater nitrate. Gridded estimates indicated a pattern of increased concentrations in intensively farmed areas, with hotspots exceeding the drinking water standard of 11.3 mg/L, though high concentrations were systematically underestimated. This approach demonstrates the first spatially continuous, data-driven estimates of median nitrate in shallow groundwater across New Zealand, offering a tool to inform water quality management and resource management. The machine learning framework can be applied in other regions where monitoring data are sparse but environmental predictors exist.