H43E-0404 1340h
Eutrophication Model Accuracy - Comparison of Calibration and Verification Performance of a Model of the Neuse River Estuary, North Carolina
A modified version of an existing two-dimensional, laterally averaged model (CE-QUAL-W2) was applied to predict water quality conditions in the lower 80-km of the Neuse River Estuary. Separate time periods were modeled for calibration and verification (model testing). The calibration time period ran from June 1997 to December 1999, while the verification time period ran from January to December 2000. During this time the estuary received two periods of unusually high inflows in early 1998 and again in September and October 1999. The latter rainfall event loaded the estuary with the equivalent of nearly two years worth of water and dissolved inorganic nitrogen in just six weeks. Overall, the level of calibration performance achieved by the model was comparable to that attained in other eutrophication model studies of eastern U.S. estuaries. The model most accurately simulated water quality constituents having a consistent spatial variation within the estuary (e.g. nitrate, salinity), and was least accurate for constituents without a consistent spatial variation (e.g. phosphate, chlorophyll-a). Calibration performance varied widely between the three algal groupings modeled (diatoms and dinoflagellates, cryptomonads and chlorophytes, cyanobacteria). Model performance during verification was comparable to the performance seen during calibration. The model's salinity prediction capabilities were somewhat better in the validation, while dissolved oxygen performance in the validation year was slightly poorer compared to calibration performance. Nutrient and chlorophyll-a performance were virtually the same between the calibration and verification exercises. As part of the TMDL analysis, an unsuccessful attempt was made to capture model error as a component of model uncertainty, but it was found that model residuals were neither unbiased nor normally distributed.
H43E-0405 1340h
Use of Selected Goodness-of-Fit Statistics to Assess the Accuracy of a Model of Henry Hagg Lake, Oregon
Assessing a model's ability to reproduce field data is a critical step in the modeling process. For any model, some method of determining goodness-of-fit to measured data is needed to aid in calibration and to evaluate model performance. Visualizations and graphical comparisons of model output are an excellent way to begin that assessment. At some point, however, model performance must be quantified. Goodness-of-fit statistics, including the mean error (ME), mean absolute error (MAE), root mean square error, and coefficient of determination, typically are used to measure model accuracy. Statistical tools such as the sign test or Wilcoxon test can be used to test for model bias. The runs test can detect phase errors in simulated time series. Each statistic is useful, but each has its limitations. None provides a complete quantification of model accuracy. In this study, a suite of goodness-of-fit statistics was applied to a model of Henry Hagg Lake in northwest Oregon. Hagg Lake is a man-made reservoir on Scoggins Creek, a tributary to the Tualatin River. Located on the west side of the Portland metropolitan area, the Tualatin Basin is home to more than 450,000 people. Stored water in Hagg Lake helps to meet the agricultural and municipal water needs of that population. Future water demands have caused water managers to plan for a potential expansion of Hagg Lake, doubling its storage to roughly 115,000 acre-feet. A model of the lake was constructed to evaluate the lake's water quality and estimate how that quality might change after raising the dam. The laterally averaged, two-dimensional, U.S. Army Corps of Engineers model CE-QUAL-W2 was used to construct the Hagg Lake model. Calibrated for the years 2000 and 2001 and confirmed with data from 2002 and 2003, modeled parameters included water temperature, ammonia, nitrate, phosphorus, algae, zooplankton, and dissolved oxygen. Several goodness-of-fit statistics were used to quantify model accuracy and bias. Model performance was judged to be excellent for water temperature (annual ME: -0.22 to 0.05 $\deg$C; annual MAE: 0.62 to 0.68 $\deg$C) and dissolved oxygen (annual ME: -0.28 to 0.18 mg/L; annual MAE: 0.43 to 0.92 mg/L), showing that the model is sufficiently accurate for future water resources planning and management.
H43E-0406 1340h
Two-Stage Automatic Calibration and Predictive Uncertainty Analysis of a Semi-distributed Watershed Model
Automatic calibration has been applied to conceptual rainfall-runoff models for more than three decades, usually to lumped models. Even when a (semi-)distributed model that allows spatial variability of parameters is calibrated using an automated process, the parameters of the model are often lumped over space so that the model is simplified as a lumped model. Our objective was to develop a two-stage routine for automatically calibrating the Soil Water Assessment Tool (SWAT, a semi-distributed watershed model) that would find the optimal values for the model parameters, preserve the spatial variability in essential parameters, and lead to a measure of the model prediction uncertainty. In the first stage of this proposed calibration scheme, a {\em global} search method, namely, the Shuffled Complex Evolution (SCE-UA) method, was employed to find the ``best'' values for the lumped model parameters. That is, in order to limit the number of the calibrated parameters, the model parameters were assumed to be {\em invariant} over different subbasins and hydrologic response units (HRU, the basic calculation unit in the SWAT model). However, in the second stage, the spatial variability of the original model parameters was restored and the number of the calibrated parameters was dramatically increased (from a few to near a hundred). Hence, a {\em local} search method, namely, a variation of Levenberg-Marquart method, was preferred to find the more distributed set of parameters using the results of the previous stage as starting values. Furthermore, in order to prevent the parameters from taking extreme values, a strategy called ``regularization'' was adopted, through which the distributed parameters were constrained to vary as little as possible from the initial values of the lumped parameters. We calibrated the stream flow in the Etowah River measured at Canton, GA (a watershed area of 1,580 km$^2$) for the years 1983-1992 and used the years 1993-2001 for validation. Calibration for daily and monthly flow produced a very good fit to the measured data. Nash and Sutcliffe coefficients for daily and monthly flow over the calibration period were 0.60 and 0.86, respectively; they were 0.61 and 0.87 respectively over the validation period. Regardless of the level of model-to-measurement fit, nonuniqueness of the optimal parameter values renders the necessity of uncertainty analysis for model prediction. The nonlinear prediction uncertainty analysis showed that cautions must be exercised when using the SWAT model to predict instantaneous peak flows. The PEST (Parameter Estimation) free software was used to conduct the two-stage automatic calibration and prediction uncertainty analysis of the SWAT model.
H43E-0407 1340h
A Statistical Model Predicting Time Series of Pesticide Load in the Sacramento River Based on Precipitation and Pesticide Use in the Sacramento River Watershed
Transport of pesticides by surface runoff during rainfall events is a major process contributing to pesticide contamination in rivers. This study presents an empirical regression model that relates pesticide loading over time in the Sacramento River with the precipitation and pesticide use in the Sacramento River watershed. The model closely simulated loading dynamics of diazinon, simazine, and diuron during 1991-1994 and 1997-2000 winter storm seasons. The coefficients of determination for regression ranged from 0.168 to 0.907, and were all significant at $<$0.001. The results of this study provide strong evidence that precipitation and pesticide use are the two major environmental variables dictating the dynamics of pesticide transport into surface water in a watershed. The capability of the statistical model to provide time-series estimates on pesticide loading in rivers is unique and may be useful for Total Maximum Daily Load (TMDL) assessment.
H43E-0408 1340h
Determining Critical Water Quality Conditions For Inorganic Nitrogen in Dry Semi-urbanized Watersheds
Traditional approaches to establishing critical water quality conditions, based on statistical analysis of low flow conditions and expressed as a recurrence interval for low-flow conditions (e.g. 7Q10), may be inappropriate for drier watersheds. The use of 7Q10 as a standard design flow assumes year-round flow, but in these watersheds 7Q10 is zero or very small. In addition, the increasing use of multiple year dynamic water quality models at daily time steps, can supercede the use of steady-state approaches. Many of these watersheds are also under increasing urbanization pressure, which accentuates the flashiness of runoff and the episodic nature of critical water quality conditions. To illustrate, we consider the conditions in the Santa Clara River, California. A statistical analysis indicates that higher inorganic nitrogen concentrations correlate strongly with low flow. However, peaks in concentrations can occur during the first storms, particularly where non-point source contribution is significant. Critical conditions can thus occur at different flow regimes depending on the relative magnitude of flow and pollutant contributions from various sources. The use of steady-state models for these dry semi-urbanized watersheds based on 7Q10 flows is thus unlikely to accurately simulate the potential for exceeding water quality objectives. Dynamic simulation of water quality is necessary, and as the recent intense storm event sampling data indicates, the models should be formulated to consider even smaller time steps. This places increasing demand on computational resources and datasets to accurately calibrate the models at this temporal resolution.
H43E-0409 1340h
Comparison of Two Watershed Models in Developing a Nutrient Management Plan for the Napa River Watershed
California has identified the Napa River Watershed as impaired due to excessive nutrient loading. A nutrient management plan, which will serve as the basis for the final TMDL to be submitted to the EPA by SFRWQCB (San Francisco Regional Water Quality Control Board), was developed for the region by using distributed parameter watershed models. Two watershed models, SWAT (Soil and Water Assessment Tools) and WARMF (Watershed Analysis Risk Management Framework), were used for this study. The two models were set up for the Napa River Watershed essentially based on same data sets, with minor adaptations based on model parameterization of processes. Although in general the models produced similar results, differences in the modeling frameworks with regards to certain non-point sources, such as atmospheric deposition, agricultural management and septic systems, resulted in local and regional differences in source strength and timing. Sensitivity analyses were conducted for the two models, to determine the robustness of modeling assumptions. Similarities and dissimilarities between two models were compared, and advantages and disadvantages of two models in terms of their future role in development of a monitoring program for the TMDL development were highlighted.
H43E-0410 1340h
A Modeling Framework for Water Quality Management in the Tongue River Watershed in Montana and Wyoming
Current drought conditions have highlighted socio-economic and resource-related conflicts in arid regions of the western United States. To address one such conflict, decision makers and resource managers have developed an analytical modeling framework to aid in the protection and utilization of water resources in the Tongue River basin in Montana and Wyoming. The modeling framework has been used to study the sources of dissolved solids in and the impact of such sources on in stream water quality. This paper describes the modeling approach, focusing specifically on methodologies developed to assess water diversions and other agricultural water management practices. Efforts to model the impacts of climate and ongoing coal bed methane development are also described. An evaluation of the predictive capability of the modeling framework, including an uncertainty analysis, is presented along with an illustration of its use in forecasting the impact of future coal bed methane development on water quality.
H43E-0411 1340h
Effects of Watershed Delineation Resolution on Water Quality Model Outputs
The first step to apply a spatially distributed water quality models, such as SWAT or WARMF, is to delineate the watershed under study at a certain spatial resolution. Model output can vary significantly for different delineations, which can represent an important source of model uncertainty. To explore the effect of delineation on model uncertainty, we applied the SWAT model in the Newport Bay watershed (California) at three delineation levels using the same Digital Elevation Model: 4-subbasin, 16-subbasin and 46-subbasin. The results indicate that stream flow is not very sensitive to delineation scheme, while simulated sediment, nutrients and pesticides concentrations depend more strongly on delineation resolution. The dependence is a function of two important classes of factors: (1) parameterization at different spatial scales; and (2) mathematical interpretation of the underlying physical processes. However, it is not clear that model output always matches observed data much better at higher resolution. The implication is that an appropriate delineation scheme should be identified to balance uncertainty and computation efficiency, and this depends on the intended use of the model.
H43E-0412 1340h
Measuring uncertainty in modeling toxic concentrations in the Niagara River
In spite of the renowned history of the Niagara River as a recreational and tourist attraction, little is known about the variability of the natural phenomena linked to the fate and transport of contaminants along the natural channel that connects Lake Erie to Lake Ontario. Previous studies of the Niagara River have focused on water quality modeling using one-dimensional deterministic models. Only recently, the work by Franceschini (2004) has evaluated probabilistically the influence on the estimated concentrations at the end of the Niagara River of few model variables and parameters with inherent variability. Several factors contribute to the uncertainty of the estimated concentrations. Toxic concentrations in the Niagara River depend on the randomly varying magnitude of the flow, the processes of decay, volatilization, and sorption which the substances undergo, the rate of contaminated suspended sediment deposition and resuspension, the variability of the input sources and seasonal schedules of water diversions for hydropower production. This paper proposes to analyze the variability of the toxic concentrations in the Niagara River with respect to the variability of selected hydraulic components (e.g. flow velocity and dispersion coefficient), upstream incoming concentrations and non-point source loadings. A comparison of the uncertainty of the results obtained during different water diversion schedules for hydropower productions will be also analyzed. The influence of climate change will be investigated with respect to temperature and precipitation variations. The uncertainty of the model results will be analyzed by Point Estimate Methods (PEMs) and specifically the Modified Rosenblueth method (Tsai and Franceschini 2004) will be used. The application of PEMs to environmental engineering problems has recently attracted some attention. Compared to other uncertainty analysis methods, PEMs require a substantially smaller computational effort, for a comparable degree of accuracy in the estimation of the first few statistical moments of a model output distribution. Furthermore, the probabilistic analysis can be used as a more rigorous method to compare the modeled results with established water quality criteria. In this study, the toxic concentrations computed at the end of the Niagara River and their estimated variability will be compared with field data measurements. The purpose of this comparison is two-fold: (a) to evaluate the accuracy of the Modified Rosenblueth method in measuring the uncertainty of toxic concentration in the Niagara River and (b) to quantify the risk of exceeding established water quality standards when such uncertainty is accounted for.
H43E-0413 1340h
Cross-Sectional And Longitudinal Uncertainty Propagation In Drinking Water Risk Assessment
Pesticide residues in drinking water can vary significantly from day to day. However, drinking water quality monitoring performed under the Safe Drinking Water Act (SDWA) at most community water systems (CWSs) is typically limited to four data points per year over a few years. Due to limited sampling, likely maximum residues may be underestimated in risk assessment. In this work, a statistical methodology is proposed to study the cross-sectional and longitudinal uncertainties in observed samples and their propagated effect in risk estimates. The methodology will be demonstrated using data from 16 CWSs across the US that have three independent databases of atrazine residue to estimate the uncertainty of risk in infants and children. The results showed that in 85% of the CWSs, chronic risks predicted with the proposed approach may be two- to four-folds higher than that predicted with the current approach, while intermediate risks may be two- to three-folds higher in 50% of the CWSs. In 12% of the CWSs, however, the proposed methodology showed a lower intermediate risk. A closed-form solution of propagated uncertainty will be developed to calculate the number of years (seasons) of water quality data and sampling frequency needed to reduce the uncertainty in risk estimates. In general, this methodology provided good insight into the importance of addressing uncertainty of observed water quality data and the need to predict likely maximum residues in risk assessment by considering propagation of uncertainties.