Impact of calibration watershed on runoff model accuracy

Hydrologic/water-quality models often need calibration to minimize differences between observed and predicted watershed responses. Often the measured data from the watershed where the models are applied are not available. Under this condition, models are calibrated using data from a different watershed having similar land use, soil, and hydrologic conditions. However, if the watershed used for model calibration is not the same as the watershed where the model is applied for hydrologic/water-quality predictions, then differences in watershed characteristics may induce errors in model output. The objective of this study was to quantify the error in model predictions when the modeled watershed is not the calibrating watershed. The Agricultural Non-Point Source (AGNPS) model was used to quantify model errors using measured data from watersheds of varying sizes within the Little Washita Basin in Oklahoma. The study indicates that error in model outputs results when a watershed different than the one of interest is used for model calibration. A significant difference in prediction error was observed between scaling-up and scaling-down conditions with mean relative error of runoff prediction being 153% for the scaling-down condition and 69% for the scaling-up condition. However, relative error of prediction showed no particular trend with the scaling ratios. A watershed having significantly higher or lower average curve number and slope condition was not a candidate watershed for calibrating the AGNPS model.