Manganese in the Northern Atlantic Coastal Plain aquifer system, eastern USA—Modeling regional occurrence with pH, redox, and machine learning

Abstract

Study region: The study was conducted in the Northern Atlantic Coastal Plain aquifer system, eastern USA, an important water supply in a densely populated region. Study focus: Manganese (Mn), an emerging health concern and common nuisance contaminant in drinking water, is mapped and modeled using the XGBoost machine learning method, predictions of pH and redox conditions from previous models, and other explanatory variables that describe the groundwater flow system and surface characteristics. Methods to address the imbalanced occurrence of elevated and low Mn concentrations are compared and used to more accurately predict concentrations of interest for human health and drinking water quality. New hydrological insights for the region: Elevated Mn concentrations were more likely in shallow groundwater, close to recharge areas and in topographically low areas where soil or unsaturated processes influence groundwater quality. Predicted concentrations greater than the health threshold of 300 micrograms per liter extended across 17 % of the surficial aquifer area, but across <1% of the areas of underlying aquifers. pH and variables related to flow-system position and near-surface processes were more important predictors than the probability of low dissolved oxygen (DO). Mapped variable influence (SHAP values) showed that both pH and DO variables were related to hydrogeologic conditions. Class weights, which improved the predictive ability for elevated Mn without altering the data, was the preferred method to address class imbalance.

Keywords

• Manganese
• Groundwater quality
• Machine learning
• Xgboost
• Class imbalance
• Classification