Skip to main content
ARS Home » Northeast Area » Beltsville, Maryland (BARC) » Beltsville Agricultural Research Center » Hydrology and Remote Sensing Laboratory » Research » Publications at this Location » Publication #387217
Research Project: Integrating Remote Sensing, Measurements and Modeling for Multi-Scale Assessment of Water Availability, Use, and Quality in Agroecosystems

Location: Hydrology and Remote Sensing Laboratory

Title: Inductive predictions of hydrologic events using a Long Short-Term Memory Network and the Soil and Water Assessment Tool

Author
item MAJESKE, N. - Indiana University
item Zhang, Xuesong
item GONG, L. - Indiana University
item ZHU, C. - Indiana University
item AZAD, A. - Indiana University

Submitted to: Environmental Modelling & Software
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 4/5/2022
Publication Date: 4/14/2022
Citation: Majeske, N., Zhang, X., Gong, L., Zhu, C., Azad, A. 2022. Inductive predictions of hydrologic events using a Long Short-Term Memory Network and the Soil and Water Assessment Tool. Environmental Modelling & Software. 152:105400. https://doi.org/10.1016/j.envsoft.2022.105400.
DOI: https://doi.org/10.1016/j.envsoft.2022.105400

Interpretive Summary: Extreme hydrologic events such as floods and droughts are among the most severe weather-related disasters that cause widespread damage in agriculture, wildlife habitats, and human properties. We examined the capability of a machine learning technique (i.e., Long Short Term Memory (LSTM) networks), to estimate soil water and streamflow spatially and temporally. We showed that LSTM networks can be trained in a fraction of the time required by the process-based Soil and Water Assessment Tool (SWAT) and achieve comparable soil moisture and streamflow prediction accuracy in the Wabash River Basin in the US Midwest. We also demonstrated that LSTM trained using observational data at multiple stream gauges can be used to estimate streamflow in other ungauged locations within the Wabash River Basin with accuracy comparable to the SWAT model. In addition, LSTM networks trained in the Wabash River Basin satisfactorily predicted streamflow in the Little River Watershed in Georgia. Overall, our results show that, at the cost of a small portion of the time required by process-based hydrologic models, LSTM can reliably predict future soil moisture and streamflow and extrapolate sparse hydrologic observations to ungauged locations

Technical Abstract: We present a machine learning method to predict hydrologic features such as streamflow and soil water from spatially and temporally varying hydrological and meteorological data. We used a temporal reduction technique to reduce computation and memory requirements and trained a Long Short-Term Memory (LSTM) network to predict soil water and streamflow over multiple watersheds. We show LSTM networks can be trained in a fraction of the time required by complex process-based and attention-based models such as Soil and Water Assessment Tool (SWAT) and GeoMAN without sacrificing accuracy. We also demonstrate that outside data - sourced from a watershed other than the target - can be used to train LSTM to comparable or even superior prediction accuracy. The success of LSTM in such spatially inductive settings shows hydrologic features can be predicted with minimal prior knowledge of the watershed in question. Finally, we make all methodologies of this work publicly available as an end-to-end software pipeline that facilitates rapid prototyping of hydrologic learners.