91´ó»ÆѼ

Informed management of drought within the western U.S. has long been conditioned on the status of seasonal snowpack, which serves as an early indicator of available water supplies and comprises the majority of warm-season runoff. In this region, projected shifts in temperature, precipitation magnitude, and precipitation phase (e.g., more precipitation falling as rain) are expected to lead to increased uncertainty in the timing and magnitude of snow and, in turn, a reduced ability to forecast drought. For a range of future projections (i.e., CMIP6 ensemble members), we quantify shifts in the timing of seasonal snow cover and the effect of declining snow resources on melt-driven runoff. We train regional-scale Long Short-Term Memory (LSTM) neural networks on a large sample of basins using daily meteorology and basin attributes (elevation, SWE/P, etc.) for four discrete time periods and report changes in the fidelity of streamflow predictions at annual and seasonal (April-July) scales. We examine drought predictability across three distinct snow regimes—coastal (n=131 basins), intermountain (n=74 basins), and continental (n=151 basins)—to identify regions with increased risk resulting from decreasing snowpack. While this analysis spans the western U.S., we recognize that drought stresses are not unique to the region. By quantifying projected changes in the ability of snow information to predict seasonal streamflow—and therefore, drought conditions—we seek to illuminate whether the relationships that dominate our current understanding of global snow-dominated water resources will remain vital in their observation, prediction, and management in the coming decades.