Open AccessAutomatic Calibration and Predictive Uncertainty Analysis of a Semidistributed Watershed Model
Author(s) -
Lin Zhulu,
Radcliffe David E.
Publication year - 2006
Publication title -
vadose zone journal
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.036
H-Index - 81
ISSN - 1539-1663
DOI - 10.2136/vzj2005.0025
Subject(s) - calibration , watershed , stage (stratigraphy) , estimator , statistics , computer science , mathematics , environmental science , hydrology (agriculture) , machine learning , geology , geotechnical engineering , paleontology
Semidistributed models are commonly calibrated manually, but software for automatic calibration is now available. We present a two‐stage routine for automatic calibration of the semidistributed watershed model Soil and Water Assessment Tool (SWAT) that finds the best values for the model parameters, preserves spatial variability in essential parameters, and leads to a measure of the model prediction uncertainty. In the first stage, a modified global Shuffled Complex Evolution (SCE‐UA) method was employed to find the “best” values for the lumped model parameters. In the second stage, the spatial variability of the original model parameters was restored and a local search method (a variant of Levenberg–Marquart method) was used to find a more distributed set of parameters using the results of the previous stage as starting values. A method called “regularization” was adopted to prevent the parameters from taking extreme values. In addition, we applied a nonlinear calibration‐constrained method to develop confidence intervals for annual and 7‐d average flow predictions. We calibrated stream flow in the Etowah River measured at Canton, GA (a watershed area of 1580 km 2 ) for the years 1983 to 1992 and used the years 1993 to 2001 for validation. The Parameter Estimator (PEST) software was used to conduct the two‐stage automatic calibration and prediction uncertainty analysis. Calibration for daily and monthly flow produced a very good fit to the measured data. Nash‐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, for the validation period. The nonlinear prediction uncertainty analysis worked well for long‐term (annual) flow in that our prediction confidence intervals included or were very near to the observed flow for most years. It did not work well for short‐term (7‐d average) flows in that the prediction confidence intervals did not include the observed flow, especially for low and high flow conditions.