Causal inference for managing groundwater nitrate levels: toward causation-based prescriptive analytics from...

This study advances groundwater nitrate management from correlation-based predictions toward causation-based prescriptions, enabling direct translation of scientific modeling outputs into actionable land use management strategies. A nationwide dataset of 3569 groundwater well samples collected across South Korea between 2023 and 2024 was integrated with high-resolution land use, elevation, and hydrogeologic unit data. The framework consisted of three phases: (i) descriptive statistical analysis to characterize spatial and hydrogeochemical patterns of groundwater chemistry, (ii) a correlation-based Random Forest (RF) model to benchmark predictive performance, and (iii) a causation-based Double Machine Learning (DML) model to quantify the causal effect of agricultural land use on groundwater nitrate concentrations. The RF model achieved strong predictive accuracy (R² = 0.53, mean absolute error = 2.26 mg/L NO₃-N) but showed a negative effect estimate when agricultural land increased by 10 percentage-points, highlighting the risk of deriving prescriptive conclusions from purely correlation-based models. In contrast, the DML model, controlling for elevation and hydrogeologic unit as confounders, estimated that a 10 percentage-point increase in agricultural land use within 200 m of a well would increase NO₃-N concentration by approximately 0.16 mg/L under identical hydrogeologic conditions. To the authors' knowledge, this represents one of the first national-scale applications of a causal modeling framework in groundwater nitrate research that directly provides prescriptive estimates for land use targets. These results demonstrate the necessity of causal inference for policy-relevant groundwater quality management and provide a transferable framework for prescribing interventions to achieve groundwater-quality goals.