DISCERNING HYDROCLIMATIC BEHAVIOR WITH A DEEP CONVOLUTIONAL RESIDUAL REGRESSIVE NEURAL NETWORK Anonymous authors Paper under double-blind review

Abstract

Water impacts the globe daily in new and familiar ways such as the ongoing western United States drought and the 2022 Pakistan flood. These events sustain uncertainty, risk, and loss forces to the global ecosystem. Better forecasting tools are mandatory to calibrate our response in an effort to mitigate such natural hazards in our watersheds and adapt to the planet's dynamic environment. Here, we present a Deep Convolutional Residual Regressive Neural Net (DCRRNN -pronounced "discern") platform for obtaining, visualizing, and analyzing the basin response of watersheds to water cycle fluxes. We examine four very large basins, simulating river response to the hydroclimatic fluxes they face. Experiments modulating the lever of time lag between remotely sensed and ground truth measurements are performed to assess the metrological limits of this forecasting device. The resultant grand mean Nash Sutcliffe and Kling Gupta efficiency values are both of greater value than 90%. Our results show that DCRRNN can become a powerful resource to simulate and forecast the impacts of hydroclimatic events as they relate to watershed response in a globally changing climate.

1. INTRODUCTION

Water is connected to and connects all living things on Earth. It is wielded to power electronic devices, enables plants, food and animals to grow, serves as the living and recreational space for many creatures big, small, young and old, and is nourishment to the human body. It has been both the subject of, platform for, and weapon of choice in numerous conflicts. Global hydraulic infrastructure is highly variable. Dirty water can be a source of disease and death. Water is branded, modified, and sold at differing levels of purity and concentration. The cost of equipment to control the flow of water is high, maintenance is frequent, and change of demand and supply is a constant source of concern. Furthermore, human activities have changed and continue to change Earth's environment. The changes are visible in both short (meteorological) and long (climatological) time scale responses (Stott, 2016) . As the temperature of our home planet increases, the amount of snow and sea ice loses volume over time (Qin et al., 2020; Min et al., 2022) , sea levels rise and swallow up once inhabited land (Tebaldi et al., 2021; Sévellec et al., 2017) , storms intensify (Karl et al., 1997) , droughts last longer Underwood (2015), floods become more severe (Milly et al., 2002; Hirabayashi et al., 2013) , animal populations go extinct (Parmesan et al., 2000) , and the availability of freshwater becomes more unreliable (Gleick & Cooley, 2021) . Concurrently, manmade Earth observation and control systems continue to improve (Crisp et al., 2020; Minzu et al., 2021) . Research, operational and pedagogical software tools for the climate sciences are interrelated by common programming interfaces and standards. In these development environments, the handling and organization of data is paramount for usability. In the United States, government supported big data systems warehousing climate data are mature. Here, we approach the topic of watershed modeling with a learned representation. We observe the connections between model output of four United States drainage basins to actual gauged in the river measurements. All basins are greater than 1M acres and one upwards of 1B. Each are substantial in size to observe how the change in runoff and subsurface flow impacts the quantity of water discharging from the 1

