A Calibration-Free Groundwater Module for Improving Predictions of Low Flows

Abstract

Introduction BackgroundShallow groundwater supplies a majority of streamflow to most watersheds (Beck et al., 2013) and is the primary source of streamflow during seasonal low flows (Smakhtin, 2001). This mobile water resource provides an essential buffer to changes in temperature, nutrients, and precipitation (Ficklin et al., 2015) and sustains evapotranspiration Yang et al., 2011) during low flow periods. Even in snow-dominated climates, the hydrologic response to snowmelt is mediated by groundwater (Enzminger et al., 2019) with important implications for the response of these systems to global warming (Tague & Grant, 2009). Because shallow groundwater contributions to low flows are disproportionately sensitive to changes in near-term (i.e., years and decades) climate signals (Hare et al., 2021), it is essential that our hydrologic models are able to predict groundwater contributions to low flows accurately. Unfortunately, accurate simulation of low flows and groundwater contributions to them has proven to be notoriously difficult in distributed, physically based land surface models (LSMs;Clark et al., 2015, Holtzman et al., 2020 as well as lumped, conceptual "bucket" models (i.e., rainfall-runoff models;Fowler et al., 2020).Early LSMs described low flows as dependent on one-dimensional drainage below a soil column (Clark et al., 2015). This "free" drainage led to well-documented inaccuracies: too-fast drainage during wet periods, underestimates of seasonal storage, and the cessation of ET during short dry periods (

Estimating Gridded Monthly Baseflow From 1981 to 2020 for the Contiguous US Using Long Short‐Term Memory (LSTM) Networks

“…LSMs provide some valuable insights into the physical processes of baseflow generation, along with the opportunity to analyze temporal baseflow trends. Nevertheless, the current generation of LSMs is flawed in predicting baseflow, low flow, and groundwater‐surface water interactions (Clark et al., 2015; Tashie et al., 2021, 2022). LSM‐derived baseflows are inherently uncertain and should be used with caution due to the following limitations: (a) The simplifying assumptions of baseflow generation processes may result in poor estimates under specific boundary conditions (Bai et al., 2016; Crow et al., 2018); (b) The absence of baseflow constraints leads to bias in determining relevant parameters (Li et al., 2019; Lo et al., 2010).…”