On the need for physical constraints in deep learning rainfall-runoff projections under climate change
NO REGISTRATION REQUIRED
Scott Steinschneider P.h.D.
Associate Professor in the Department of Biological and Environmental Engineering at Cornell University
Deep learning (DL) rainfall-runoff models have recently emerged as state-of-the-science tools for hydrologic prediction that outperform conventional, process-based models in a range of applications. However, it remains unclear whether DL models can produce physically plausible projections of streamflow under significant amounts of climate change. We investigate this question here, focusing specifically on modeled responses to increases in temperature and potential evapotranspiration (PET). Previous research has shown that increases in temperature should lead to declines in streamflow absent any changes to precipitation, and also that temperature-based methods to estimate PET lead to overestimates of water loss in rainfall-runoff models under warming, as compared to energy budget-based PET methods. We leverage these insights to test the physical plausibility of DL-based streamflow projections under warming in two case studies in California and the Great Lakes basin, as well as across the contiguous United States. We focus on regionally trained Long Short-Term Memory networks (LSTMs) and several physics-informed variants, and use multiple process-based rainfall-runoff models as benchmarks. The primary results of this work demonstrate that although DL rainfall-runoff models provide the most robust historical out-of- sample predictions of streamflow under space-time cross-validation, they struggle in some cases to extrapolate plausible hydrologic responses under significant warming, both in terms of the direction and magnitude of streamflow change. The primary driver of implausible streamflow projections is related to historical associations between temperature and PET that are geographically pervasive but unlikely to remain stationarity under climate change. Our results also demonstrate that physical constraints regarding model architecture and input variables can promote physical realism in DL-based hydrologic projections under future warming.
Scott Steinschneider is an Associate Professor in the Department of Biological and Environmental Engineering at Cornell University. His research program enhances the sustainability of water systems through innovations in climate risk management to support decision- making under uncertainty. He advances this program through three inter-related themes, including: 1) assessing the variability, predictability, and change of hydroclimate processes; 2) quantifying the risks these processes pose to water services; and 3) identifying robust adaptation strategies that can mitigate these risks. Dr. Steinschneider's work has focused on water systems across the United States and globally and has been sponsored by the U.S. Army Corps of Engineers, National Atmospheric and Oceanic Administration, New York Sea Grant, National Science Foundation, and US Department of Agriculture. Dr. Steinschneider earned his B.A. in Mathematics from Tufts University and his M.S. and Ph.D. in Civil and Environmental Engineering from the University of Massachusetts, Amherst. Prior to arriving at Cornell, he was a postdoctoral research fellow at Columbia University.
