Use of multiple modules and Bayesian model averaging to assess structural uncertainty of catchment-scale wetland modeling in a coastal plain landscape

Authors: LEE, S. - University Of Maryland; YEN, H. - Texas A&M University; YEO, I.Y. - University Of Newcastle; Moglen, Glenn; RABENHORST, M. - University Of Maryland; McCarty, Gregory

Submitted to: Journal of Hydrology
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 1/3/2020
Publication Date: 1/7/2020
Citation: Lee, S., Yen, H., Yeo, I.Y., Moglen, G.E., Rabenhorst, M., McCarty, G.W. 2020. Use of multiple modules and Bayesian model averaging to assess structural uncertainty of catchment-scale wetland modeling in a coastal plain landscape. Journal of Hydrology. https://doi.org/10.1016/j.jhydrol.2020.124544.
DOI: https://doi.org/10.1016/j.jhydrol.2020.124544

Interpretive Summary: Uncertainty from model predictions comes from several sources. Understanding the role and magnitude from these different sources is important so that predictions can be as reliable and accurate as possible. Problems come from model structures that neglect physical processes and the inherent spatial variabiilty of the natural world. This study examines all of these concerns in the context of wetland hydrologic modeling during flood and drought with several different complexities of model structure. This study identifies the model structure that performs best and the reasons for this strong performance. Following on the results presented here, future wetland modeling efforts are likely to be more accurate so that conclusions drawn about the role of wetlands (or wetland loss) in moderating larger scale conditions at the hydrologic extremes of flood and drought are well-informed.

Technical Abstract: Wetlands are an important landscape feature that provide multiple ecosystem services. Hydrologic models are used as a tool to estimate measurable benefits from wetlands, but the magnitude of structural uncertainty over catchment-level wetland modeling remains largely unknown. This study used the Soil and Water Assessment Tool (SWAT) to assess wetland modeling structural uncertainty on streamflow prediction and wetland loss impact quantification. Two wetland modules, the riparian wetland module (RWM) and the geographically isolated wetland module (GIWM), were employed in SWAT to create three wetland modeling structures: 1) use of the default SWAT wetland module for riparian wetlands (RWs) and geographically isolated wetlands (GIWs, SS method), 2) use of the RWM for RWs and the SWAT module for GIWs (RS method), and 3) use of RWM for RWs and GIWM for GIWs (RG method). Bayesian Model Averaging (BMA) was used to generate an ensemble of the three model structures. To reduce parameter uncertainty on model outputs, we used a parameter set that was mutually acceptable for all three wetland modeling structures. Results showed that the RG method provided the best streamflow simulation results and its prediction of low-flow conditions was notably more accurate relative to the other methods. The RG method also captured wetland water storage functions, such as storing excessive water under wet conditions and supporting low-flow conditions under drier conditions. In contrast, the two other methods did not predict drought mitigation by wetlands. Relative to the RG method, the BMA results indicated less accurate model prediction and inadequate representation of wetland loss impacts due to poor predictions by the SS and RS methods. The superior performance of the RG method was attributed to an enhanced model structure relative to the other methods. These results suggest that wetland modeling development should be toward an explicit spatial characterization of the wetland.