Keywords
uncertainty, catchment, water quality, modelling
Start Date
1-7-2004 12:00 AM
Abstract
Sources of Uncertainty Global Assessment using Split SamplES (SUNGLASSES) is a method for assessing model global uncertainty to aid in the development of integrated models. The method is complementary to the commonly investigated input and parameter uncertainty, as it accounts for errors that may arise due to unknown or unassessable sources of uncertainty, such as model hypothesis errors, simplifications, scaling effects or the lack of the observation period to represent long-term variability and fluctuations in the system. Such sources are typically dominant for most environmental models and they undermine the reliability of environmental models. The SUNGLASSES algorithm directly estimates the overall predictive uncertainty without identifying or quantifying the underlying sources of uncertainties. The method uses the split sample approach to generate an estimate of model output uncertainty by selecting a threshold below which model simulations are determined to be acceptable. Where this methodology differs from other methods that use a threshold, is that the threshold is determined by evaluating the confidence bounds on model outputs during an evaluation time period of data that was not used to initially calibrate the model and generate parameter estimates. Where parameter uncertainty is often assessed using some goodness-of-fit criterion such as the mean squared errors, SUNGLASSES focuses on a criterion that evaluates the correctness of the model output values to be used directly in decision making, such as total mass balance assessments or violations of standards as imposed by legislation. The described method is applied to the integrated water quality modelling tool, SWAT2003, applied to Honey Creek, a tributary of the Sandusky catchment in Ohio. Water flow and sediment loads are analysed. The incorporation of the split sample approach in the methodology produces a reasonable error bound that captures most of the observations during both the initial calibration period and during the evaluation period.
Dealing with unidentifiable sources of uncertainty within environmental models
Sources of Uncertainty Global Assessment using Split SamplES (SUNGLASSES) is a method for assessing model global uncertainty to aid in the development of integrated models. The method is complementary to the commonly investigated input and parameter uncertainty, as it accounts for errors that may arise due to unknown or unassessable sources of uncertainty, such as model hypothesis errors, simplifications, scaling effects or the lack of the observation period to represent long-term variability and fluctuations in the system. Such sources are typically dominant for most environmental models and they undermine the reliability of environmental models. The SUNGLASSES algorithm directly estimates the overall predictive uncertainty without identifying or quantifying the underlying sources of uncertainties. The method uses the split sample approach to generate an estimate of model output uncertainty by selecting a threshold below which model simulations are determined to be acceptable. Where this methodology differs from other methods that use a threshold, is that the threshold is determined by evaluating the confidence bounds on model outputs during an evaluation time period of data that was not used to initially calibrate the model and generate parameter estimates. Where parameter uncertainty is often assessed using some goodness-of-fit criterion such as the mean squared errors, SUNGLASSES focuses on a criterion that evaluates the correctness of the model output values to be used directly in decision making, such as total mass balance assessments or violations of standards as imposed by legislation. The described method is applied to the integrated water quality modelling tool, SWAT2003, applied to Honey Creek, a tributary of the Sandusky catchment in Ohio. Water flow and sediment loads are analysed. The incorporation of the split sample approach in the methodology produces a reasonable error bound that captures most of the observations during both the initial calibration period and during the evaluation period.
