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

• Published in: Water resources research Vol. 58; no. 8
• Main Authors: Xie, Jiaxin, Liu, Xiaomang, Tian, Wei, Wang, Kaiwen, Bai, Peng, Liu, Changming
• Format: Journal Article
• Language: English
• Published: Washington John Wiley & Sons, Inc 01-08-2022
• Subjects: Base flow; baseflow; Basins; Complexity; Datasets; Deep learning; Discharge measurement; Drought; Estimation; Fish; Food production; Groundwater; Groundwater recharge; Hydrologic data; long short‐term memory; Median (statistics); Monthly; Networks; Properties; Resolution; Spatial discrimination; Spatial resolution; Stream discharge; Stream flow; Streams; Time series; time series simulation; Training; Trends; Water resources; Water resources management; Water supply; United States > US
• ISSN: 0043-1397; 1944-7973
• DOI: 10.1029/2021WR031663

Accurate baseflow estimation is essential for ecological protection and water resources management. Past studies have used environmental predictors to extend baseflow from gauged basins to ungauged basins, publishing several regional or global datasets on mean annual baseflow. However, time series datasets of baseflow are still lacking due to the complexity of baseflow generation processes. Here, we developed a monthly baseflow data set using a Deep learning model called the long short‐term memory (LSTM) networks. To better train the networks across basins, we compared the standard LSTM architecture using 8 time series as inputs with four variant architectures using 16 additional static properties as inputs. Dividing the contiguous United States into nine ecoregions, we used baseflow calculated from 1,604 gauged basins as training targets to calibrate the five LSTM architectures for each ecoregion separately. Results show that three variant architectures (Joint, Front, and Entity‐Aware‐LSTM) perform better than the standard LSTM, with median Kling‐Gupta Efficiencies across basins greater than 0.85. Based on Front LSTM, the monthly baseflow data set with 0.25° spatial resolution across the contiguous United States from 1981 to 2020 was obtained. Potential applications of the data set include analyzing baseflow trends under global change and estimating large‐scale groundwater recharge.
Plain Language Summary
After precipitation has ceased for some time, upstream banks slowly release groundwater into streams through delayed pathways. This type of flow is known as baseflow. Baseflow is important because it maintains streamflow during prolonged droughts, providing water supplies for food production and fish habitat. Several methods have been developed to estimate baseflow from streamflow. However, due to the complex processes of baseflow generation, it is challenging to estimate baseflow for a basin where streamflow is unavailable. This study introduced a deep learning model called Long Short‐Term Memory (LSTM) network to bridge the gap. To better train across basins, we compared the standard LSTM with four variant architectures that incorporate additional static properties as inputs. Results show that three variants (Joint, Front, and Entity‐Aware‐LSTM) perform better than the standard LSTM, with median Kling‐Gupta Efficiencies across basins greater than 0.85. Training the Front LSTM separately in nine ecoregions of the contiguous United States, we identified the relationship between 24 spatiotemporal features and the baseflow calculated from streamflow in 1,604 basins, thereby deriving a monthly baseflow data set at 0.25° spatial resolution across the contiguous United States. This data set provides an opportunity to analyze large‐scale baseflow trends and is beneficial for groundwater estimation.
Key Points
We develop a monthly baseflow data set at 0.25° resolution using Deep learning networks (Long Short‐Term Memory, LSTM) with 8 dynamic time series as inputs
Three LSTM variants using 16 static properties as additional inputs outperform the standard LSTM, with median Kling‐Gupta Efficiencies across 1,604 basins greater than 0.85
Dynamic inputs interpret 77.8% in baseflow simulation, with P and potential evapotranspiration contributing the most, while slope dominates the static inputs