Exploring Climate-Driven agricultural water shortages in a Snow-Fed basin using a water allocation model and Machine learning

Abstract

Water allocation models (WAM) can capture complex feedbacks between water infrastructure, water supply, water demand, and water rights structure. They are an important tool for water managers to reduce conflict and aid in future water resource planning. Future streamflow reductions and shifts in timing are anticipated in snow-dominated watersheds of the western United States due to projected reductions in snow water storage. These basins tend to be water-limited and rely on groundwater to support agricultural demand. Water managers in the snow-fed, agricultural Walker River Basin (WRB) in California and Nevada use a dynamically coupled WAM and physically based groundwater flow model as a decision support tool. Large computation time is required to simulate groundwater dynamics, limiting the number of model scenarios that can feasibly be run to test alternative management scenarios under climate change. Machine learning (ML) techniques have been used to capture complex nonlinear hydrologic dynamics in other contexts, but to our knowledge there have not been efforts to use ML to increase computational efficiency in modeling groundwater responses to agricultural water use. We use extreme gradient boosting to replace the groundwater sub-model for the WRB decision support tool, resulting in accurate estimates of reservoir storage (monthly error almost always < 5% of reservoir capacity), streamflow (total outflow errors from −3% to 20% depending on scenario) and agricultural water shortages (-12% to 18% error in total shortage depending on scenario). For comparison, we found that omitting groundwater modeling from the WAM results in a larger overprediction of outflows (2 to 58%) and underprediction in agricultural water shortages (0 to 41%) across a range of scenarios. We use this validated ML method to explore the sensitivity of agricultural water availability to a range of streamflow volumes and streamflow timings. Results indicate that agricultural water shortages are more sensitive to annual streamflow volume in comparison to shifts in streamflow timing. Our model shows that reservoirs can buffer the impact of altered flow timing on agricultural production, except under extreme conditions in which all precipitation falls as rain (no snow storage) and streamflow peaks in the winter. Future work will explore these thresholds in more detail. While the results here are site-specific, the methodology for coupling a WAM with a ML representation of surface–groundwater interactions can be applied to a wide variety of water resource modeling applications. These results also provide a stark example of the importance of snowpack storage and groundwater management to western US water infrastructure.