A water-balance-guided hybrid machine-learning framework for quantifying recharge to assess groundwater depletion in the Southeastern USA

Study region Southeastern United States Study focus Groundwater recharge is an important hydrological flux; however, recharge estimation methods used in the literature estimate different types of recharge fluxes. The focus of the current study is to elucidate the distinction between two common recharge estimates, namely water-balance (WB) recharge, which represents water leaking below the root zone, and water-table fluctuation (WTF) recharge, which represents the net amount of recharge reaching the groundwater table. We developed a water-balance-guided hybrid machine-learning framework that first estimates WB recharge and uses it as a predictor to constrain net recharge predictions. Several machine-learning models were evaluated, and a multilayer perceptron (MLP) was selected to be the best-performing model. The new ML-model-derived data are mapped at the HUC12 scale across the Southeastern US. The modeled recharge was then used to quantify groundwater depletion (GWD) levels as the imbalance between recharge and total fresh groundwater withdrawals. New hydrological insights Results show that the net recharge for the region generally ranges from about 10–15% of average annual local rainfall for most of the catchments. Furthermore, estimated GWD levels exhibited noticeable spatiotemporal variability, with cumulative depletion reaching approximately 31 km³ during 2008–2020 with an average annual depletion rate of about 2.5 km³ /year. This study highlights the importance of distinguishing different types of recharge estimates and also demonstrates the need for using net recharge in GWD assessments.