Improving probabilistic monthly water quantity and quality predictions using a simplified residual-based modeling approach

Authors: GUO, TIAN - Purdue University; LIU, YAOZE - State University Of New York (SUNY); SHAO, GANG - Purdue University; ENGEL, BERNARD - Purdue University; SHARMA, ASHISH - University Of New South Wales; MARSHALL, LUCY - University Of New South Wales; Flanagan, Dennis; CIBIN, RAJ - Pennsylvania State University; WALLACE, CARLINGTON - Interstate Commission On The Potomac River Basin; ZHAO, KAIGUANG - The Ohio State University; REM, DONGYANG - Purdue University; VERA MERCADO, JOHANN - Purdue University; ABOELNOUR, MOHAMED - University Of Notre Dame

Submitted to: Environmental Modelling & Software
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 8/10/2022
Publication Date: 8/16/2022
Citation: Guo, T., Liu, Y., Shao, G., Engel, B.A., Sharma, A., Marshall, L.A., Flanagan, D.C., Cibin, R., Wallace, C.W., Zhao, K., Rem, D., Vera Mercado, J., Aboelnour, M.A. 2022. Improving probabilistic monthly water quantity and quality predictions using a simplified residual-based modeling approach. Environmental Modelling & Software. 156: Article 105499. https://doi.org/10.1016/j.envsoft.2022.105499.
DOI: https://doi.org/10.1016/j.envsoft.2022.105499

Interpretive Summary: Computer simulation models are often used to estimate occurrences and amounts of surface runoff, subsurface drainage, sediment losses, and pollutant losses (nitrogen, phosphorus, pesticides, etc.). The model developers attempt to consider observed data and natural processes as much as possible, but since any model is just a simplified representation of nature, any prediction has some level of error or uncertainty associated with it. Often model predictions are compared to observed data from plots, fields, or watersheds, to calculate statistics on the level of agreement between the model results and the observations. This study examined different approaches to quantify the levels of uncertainty. We used information from 18 locations across the U.S., and two USDA models, the Soil and Water Assessment Tool (SWAT) and the Water Erosion Prediction Project (WEPP) models. Simplified approaches to generate probabilistic hydrologic and water quality predictions and evaluate residual error model performance for nine different residual error transformation schemes were applied. One transformation scheme group provided better reliability, precision, and volumetric bias metrics than the other two residual error scheme groups for all case studies. Uncertainty prediction from the best transformation schemes was consistent between model calibration and validation, between hydrologic and water quality simulations, and between a study area and adjacent study areas. Other scientists, university faculty, and students are impacted by these results. Results from this study could improve approaches to estimate natural resource model uncertainty estimation in hydrologic and water quality simulations at various locations.

Technical Abstract: Hydrologic and water quality model simulations have guided water quality assessment and water resource management. However, uncertainties between simulated and observed nutrient loads need to be quantified through statistical methods such as residual error analysis. This study aims to simulate monthly hydrology and water quality residual errors to improve probabilistic predictions in diverse watersheds/fields. The calibrated streamflow, baseflow, subsurface drainage flow, surface runoff, sediment, nitrogen, nitrate, and phosphorus results from two hydrologic models and the observed data at 18 watersheds/fields across the U.S. were used for residual error analysis. A simplified approach (Least Square-Method of Moments method) was used to generate probabilistic hydrological and water quality predictions and evaluate residual error models’ performance under nine residual error transformation schemes. A transformation scheme group (BC0.5_A0, BC0.5_A0.0001, and BC0.5_A0.1) could provide better reliability, precision, and volumetric bias metrics than the other two residual error scheme groups for all case studies. The performance metrics with three BC_0.2 schemes were similar for hydrologic and water quality predictive uncertainties across all case studies. The tradeoffs in the performance metrics for a single variable predictive uncertainty in a single study watershed became clearer than those for all hydrologic or water quality cases. Compared to a single realization of simulations, the ensemble average of hydrologic and water quality simulations can reduce predictive uncertainty, especially for large watersheds. The predictive uncertainty performance from the best transformation schemes was consistent between calibration and validation, between hydrologic simulation and water quality simulation, between a realization of simulation and other realizations, and between a study area and adjacent study areas. This study recommends various opportunities via residual error scheme selection, data monitoring improvement, and hydrologic model enhancement to robust hydrologic and water quality predictive uncertainties. The results could improve the quantification of the predictive uncertainty of hydrologic and water quality simulations in various locations and guide probabilistic prediction enhancement.