Authors: John T. Kemper, Kristen L. Underwood, Andrew W. Schroth – University of Vermont; Scott D. Hamshaw, James B. Shanley – U.S. Geological Survey
Title: Forecasting water quality from National Water Model outputs at actionable scales
Abstract: As water quality monitoring continues to expand globally, freely accessible, process-based forecasting models have improved anticipation of flow events, enabling data-driven models of water quality to be coupled to streamflow projections to forecast water quality at many locations. Here we develop machine learning (ML) models of water quality in two heterogeneous watersheds in the northeastern USA and combine them with streamflow forecasts from the U.S. National Water Model (NWM) to develop forecasts of water quality concentration and flux at actionable spatial and temporal scales. Each test case involves constituents of primary concern within their respective basin: turbidity loading in Esopus Creek, a source watershed of the New York City reservoir system, and phosphorus loading in the Lake Champlain basin of Vermont and New York. For Esopus Creek, interpretation of feature weights within ML models indicates that erosional processes in specific sub-basins elevate turbidity for a given discharge following large storm events and overall influence catchment-scale turbidity behavior. ML models regularly offer improvements for turbidity prediction over linear methods, emphasizing their capabilities as forecasting tools. In the Lake Champlain basin, model skill appears to be inversely related to hydrologic dynamicity, suggesting that water quality in basins where high flow conditions are marked departures from average behavior is harder to predict. Additionally, watershed-specific models offer improvements over a lumped approach as basins get wetter and less agricultural. The forecasting approach presented here, with flexible, efficient ML models that are readily interpretable, can potentially inform efforts to design or improve models of water quality in basins throughout the globe. Taken together, these results clearly demonstrate the broad utility of the NWM as a water quality forecasting tool, which, when coupled with robust long-term or high frequency water quality monitoring data, can provide water quality forecasts at stakeholder relevant timescales across the country.