Downscaling water storage: forcasting floods in the Upper Mississppi Basin utlizing downscaled GRACE data

Abstract

Flooding is one of the world's most financially devastating natural hazards, and understanding storage-discharge relations can improve existing flood forecasts. The Gravity Recovery and Climate Experiment (GRACE) mission, used for monitoring water storage since 2003 and has been tested for storage-discharge relations since 2009, remains underdeveloped in its application to streamflow predictions because the current resolution of GRACE total water storage (GRACETWS) anomalies remains coarse (1o x 1o). Furthermore, the current downscaled GRACETWS cannot reflect the geography, hydrological characteristics, and localized level of the water balance of watersheds; the dissertation seeks to address this challenge through three interconnected experiments. The first part of the dissertation assessed whether a machine learning (ML) algorithm, like the Long Short-Term Memory (LSTM) network, could improve its streamflow prediction if synthetic total water storage (ΔS), simulated by the physically-based Hydrological Simulation Program-Fortran (HSPF) in the Rum River Watershed and the Wild Rice River Watershed, was used as a dynamic input. The second part analyzed storage-discharge hysteresis, the non-linear relation caused by heterogeneous geographic characteristics of the watershed. Virtual experiments, based on the water retention curve (WRC) and the nonlinear Preisach model, are used to test the characteristics of the storage-discharge hysteresis on conceptual models created by the combinations of linear reservoirs and the HSPF model for the Rum River Watershed. Finally, the third part assesses whether we can apply the storage-discharge relation and ML algorithm’s capacity to predict the spatial water storage distribution within the GRACETWS. It tests downscaling GRACETWS to a smaller, 1744 sub-watershed using the Long Short-Term Memory (LSTM) network between 2009 and 2015, located in the Upper Mississippi River Watershed within the boundary of Minnesota and assesses the improvement of streamflow prediction by inputting the downscaled GRACETWS for the universal LSTM rainfall-runoff model used in the first part of the experiment. The first part shows that the LSTM can improve its streamflow prediction by including synthetic water storage value, but the improvement accuracy varies based on the correlation between storage and discharge. The second experiment indicated that the storage-discharge hysteresis behaviors can be analyzed like a Preisach model, and the HSPF model can help us analyze how nonlinearity and heterogeneity contribute to storage-discharge hysteresis. Finally, the downscaling proves that the downscaled total water storage can potentially match the well-depth of the local location and can be used for improving streamflow, but it requires future research on improving data assimilation and systematic study on downscaled GRACETWS and selected scale of watershed outflow.