Mapping Denitrification Potential During Infiltration with Machine Learning Informed by Field and Laboratory Information

Abstract:

Managed aquifer recharge (MAR) can increase groundwater storage and, under some conditions, improve groundwater quality simultaneously. We combine observations from laboratory tests, field experiments, and operational MAR facilities at four MAR locations in the Pajaro Valley in central coastal California, USA to develop a predictive model of potential nitrate removal (NRp) during infiltration. We combine data on soil and fluid properties with conditions during infiltration, and compare three modeling approaches: multiple linear regression, random forests, and boosted regression trees. Our preferred model uses boosted regression trees based on four predictor variables: total soil carbon, soil clay fraction, fluid residence time, and initial nitrate concentration. We apply this model to the heterogeneous and mixed-use landscape of the Pajaro Valley, to assess where water quality improvement might be achieved during MAR. We find that areas of high NRp are more common on floodplains and riparian areas, and urban areas tend to have higher NRp than do forested or agriculturally developed areas. We combine maps of NRp with modeled hillslope runoff (from independent stormwater routing calculations) simulated under varying climate scenarios using downscaled meteorological data. This links high nutrient loads carried by runoff to the nutrient cycling capacity of ambient soils, where simultaneous benefits may be achieved in meeting water supply and quality goals. To analyze the uncertainty in model results, we compare uncertainties associated with our input datasets to those introduced from the modeling framework. Finally, we apply the NRp model to other agricultural areas across California to illustrate the utility and flexibility of this approach, which could help guide decisions in resource management and identify promising MAR sites. NRp simulations also show where more and better experimental data might be especially useful, leading to improved understanding of biogeochemical cycling capacity at large spatial scales.