HR: 08:00h
AN: H11B-01    [PDF]
TI: Assessing Model Structural Uncertainty Using a Split Sample Approach for a Distributed Water Quality Model
AU: * Meixner, T
EM: tmeixner@mail.ucr.edu
AF: Dept. of Environmental Sciences, University of California, Room 2217 Geology, Riverside, CA 92507 United States
AU: van Griensven, A
EM: annvg@mail.ucr.edu
AF: Dept. of Environmental Sciences, University of California, Room 2217 Geology, Riverside, CA 92507 United States
AB: A method for assessing model structural uncertainty as opposed to the more commonly investigated parameter uncertainty is presented that should aid in the development of improved water quality models. Elsewhere (see van Griensven and Meixner abstract, this session) we have developed a methodology (ParaSol) to estimate model parameter uncertainty. Uncertainty is typically estimated with a specific time period of data. However from experience with model calibration problems we know that we need to employ split sample and other evaluation tests to estimate the confidence we should have in our models and our methods. Evaluation tests generally give us qualitative data about confidence in our models. Here we propose a method that uses the split sample approach to generate a quantitative estimate of model structural uncertainty. The Sources of Uncertainty Global Assessment using Split SamplES (SUNGLASSES) method is designed to assess predictive uncertainty that is not captured by parameter or physical input uncertainty. We assume that this additional uncertainty represents model structural error in how the model represents the physical, chemical, and biological processes incorporated into water quality models. This method operates by selecting a threshold for a sample statistic (bias in our case), when the sample statistic for a model simulation is below the threshold the simulation is acceptable. Where this methodology differs from others is that the threshold is determined by evaluating whether the chosen threshold will capture simulations during an evaluation time period (hence split sample) that was not used to initially calibrate the model and generate parameter estimates. Most existing methods rely solely on sample statistics during a calibration period. The new method thus captures an element of predictive error that originates in the structural conception of the processes controlling water quality. The described method is applied on a Soil Water Assessment Tool (SWAT) model of Honey Creek, a tributary of the Sandusky catchment in Ohio. Water flow and sediment loads are analyzed. The results show the minor importance of model parameter uncertainty assessed by ParaSol in view of the total model uncertainty that was assessed by SUNGLASSES.
DE: 1869 Stochastic processes
DE: 1871 Surface water quality
DE: 1894 Instruments and techniques
DE: 3210 Modeling
DE: 3260 Inverse theory
SC: Hydrology [H]
MN: 2003 Fall Meeting