HR: 1340h
AN: H43E-0406 [Abstracts]
TI: Two-Stage Automatic Calibration and Predictive Uncertainty Analysis of a Semi-distributed Watershed
Model
AU: * Lin, Z
EM: linzhulu@uga.edu
AF: Department of Crop and Soil Sciences,
The University of Georgia, 3102B Miller Plant Sciences Building, Athens, GA 30602
United States
AU: Radcliffe, D E
EM: dradclif@uga.edu
AF: Department of Crop and Soil Sciences,
The University of Georgia, 3102B Miller Plant Sciences Building, Athens, GA 30602
United States
AU: Doherty, J
EM: jdoherty@gil.com.au
AF: Department of Civil Engineering, University of Queensland, St. Lucia, Queensland, 4072
Australia
AB:
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.
DE: 1860 Runoff and streamflow
DE: 1869 Stochastic processes
DE: 1871 Surface water quality
SC: Hydrology [H]
MN: 2004 AGU Fall Meeting