Publications

Fleming, SW; Rittger, K; Taglialatela, CMO; Graczyk, I (2024). Leveraging Next-Generation Satellite Remote Sensing-Based Snow Data to Improve Seasonal Water Supply Predictions in a Practical Machine Learning-Driven River Forecast System. WATER RESOURCES RESEARCH, 60(4), e2023WR035785.

Abstract
Seasonal predictions of spring-summer river flow volume (water supply forecasts, WSFs) are foundational to western US water management. We test a new space-based remote sensing product, spatially and temporally complete (STC) MODSCAG fractional snow-covered area (fSCA), as input for the Natural Resources Conservation Service (NRCS) operational US West-wide WSF system. fSCA data were considered alongside traditional SNOTEL predictors, in both statistical and AI-based NRCS operational hydrologic models, throughout the forecast season, in four test watersheds (Walker, Wind, Piedra, and Gila Rivers in California, Wyoming, Colorado, and New Mexico). Outcomes from over 200 WSF models suggest fSCA-enabled accuracy gains are most consistent and explainable for short-lead, late-season forecasts (roughly 10%-25% improvements, typically), which in operational practice can be challenging as snowlines rise above in situ measurement sites. Gains are roughly proportional to how thoroughly spring-summer runoff is dominated by snowmelt, and how poorly in situ networks monitor late-season snowpack. fSCA also improved accuracy for long-lead, early-season forecasts, which are similarly problematic in WSF practice, but not for WSFs issued around the time of peak snow accumulation, when in situ measurements reasonably characterize mountain snowpack available for upcoming spring-summer snowmelt. The AI-based hydrologic model generally outperformed the statistical model and, in some cases, better-capitalized on satellite remote sensing. Additionally, preliminary analyses suggest reasonable WSF skill in many cases using fSCA as the sole predictor, potentially useful in sparsely monitored regions; and that combining satellite and in situ products in data-driven hydrologic models using genetic algorithm-based predictor selection could help guide new SNOTEL site selection. Western US operational water supply forecasts (WSFs) are predictions, typically issued at the start of every month from January through spring, of upcoming spring-summer flow volume for a given point on a river, performed by service-delivery organizations having strict accountabilities to end users around reliably delivering this information. WSFs use mathematical models of watershed hydrology that heavily leverage on-the-ground data on winter mountain snowpack, source of much of the spring-summer runoff. Past research shows that snow measurements from air and space can improve WSF accuracy, but operational WSF models often don't directly ingest such data for practical reasons. Here, we consider new, NASA satellite-derived snow data that is uniquely suited to WSF operations, partly because coverage in space and time is gap-free, and test it in the largest stand-alone operational WSF system in the American West, run by the US Department of Agriculture. Overall, outcomes demonstrate the benefits of this satellite data in statistical and AI-based hydrologic models used in standard WSF applications in the western US, and guide scientists and engineers around when and where to use those data. Such WSF improvements will be critical to successful water management going forward in this increasingly water-stressed region. Improvements to operational water supply forecasts (WSFs), based heavily on mountain snow data, are critical to western US water management We test a new satellite remote sensing snow product, having no spatial or temporal coverage gaps, for ability to improve USDA WSF models Results argue for operational implementation and give practical guidance around when and where such data may provide the most benefit

DOI:
10.1029/2023WR035785

ISSN:
1944-7973