Estimation of base and surface flow using deep neural networks and a hydrologic model in two watersheds of the Chesapeake Bay

Authors: LEE, J. - Ulsan National Institute Of Science And Technology (UNIST); ABBAS, A. - Ulsan National Institute Of Science And Technology (UNIST); McCarty, Gregory; Zhang, Xuesong; LEE, S. - University Of Seoul; CHO, K.H - Ulsan National Institute Of Science And Technology (UNIST)

Submitted to: Journal of Hydrology
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 12/13/2022
Publication Date: 12/24/2022
Citation: Lee, J., Abbas, A., McCarty, G.W., Zhang, X., Lee, S., Cho, K. 2022. Estimation of base and surface flow using deep neural networks and a hydrologic model in two watersheds of the Chesapeake Bay. Journal of Hydrology. 617. Article 128916. https://doi.org/10.1016/j.jhydrol.2022.128916.
DOI: https://doi.org/10.1016/j.jhydrol.2022.128916

Interpretive Summary: The ability to accurately separate total stream flow into base and storm flows provides valuable information to support decision making in controlling floods, operating water resources, and mitigating water quality degradation. Various watershed models, like SWAT, have been used to perform flow separation but their accuracy is generally reduced by data input requirements and model structure limitations. In this study, we tested the utility of artificial intelligence (deep learning) methods for generating more accurate stream flow separation. Our results demonstrated that the LSTM deep learning model generated superior flow predictions when compared to traditional models. Use of improved stream flow predictions will enhance capability to manage the water quality and quantity concerns in agricultural watersheds.

Technical Abstract: Flow simulation provides critical information to support water resource management and prevent water quality degradation. Here, we propose a deep-learning approach that uses both sub-basin characteristics and long-term meteorological data to drive the long-short-term memory (LSTM) model. We compared the sub-basin-level LSTM model with the process-based Soil and Water Assessment Tool (SWAT) model and a basin-level LSTM for simulating base, surface, and total flow in the Tuckahoe Creek watershed (TCW) and Greensboro watershed (GW) located in the Northeastern United States. In general, the sub-basin-level LSTM model performed pronouncedly better for total flow simulation for the 20-year simulation period. (Nash-Sutcliffe Efficiency [NSE ]=0.48 in TCW and NSE=0.60 in GW) than the SWT model (NSE=0.40 in TCW and NSE=0.44 in GW) and the basin-level LSTM model (NSE=0.44 in TCW and NSE=0.45). Furthermore, the sub-basin-level model can accurately predict the peaks in total and base flow in both wet and dry years. We also used the model interpretation method, Shapely Additive exPlanations (SHAP) to analyze the importance of input features to sub-basin-level LSTM model and found minimum temperature was the most influential feature on the flow simulation. Overall, the LSTM model with sub-basin-level characterization of a watershed is recommended to provide long-term simulation of flow to support water resources management, particularly when time and resources are limited