Using a global sensitivity analysis to estimate the appropriate length of calibration period in the presence of high hydrological model uncertainty

Abstract

Generally, reliable simulation results are obtained by calibrating and validating the parameters of a hydrological model through a split-sample test. To obtain adequate and stable calibration and validation results of a hydrological model, calibration data with sufficient length should be used; therefore, a study on the estimation of appropriate calibration data length is required. In this study, the appropriate calibration data length was estimated for three hydrological models (GR4J, IHACRES, and Sacramento models) with varying complexities using the Sobol global sensitivity analysis method. As a result of analyzing the appropriate calibration data length in three hydrological models for three dam catchments in Korea, a relatively stable simulation result could be derived using more than eight years of calibration data length. In addition, it was confirmed that the appropriate length of the calibration data increased as the size of the catchment decreased. In the case of the Sacramento model with the largest number of parameters, the variability of the optimum parameter values was high, even if the calibration period increased. Therefore, the variability size of the optimum parameter values is an improper scale for estimating the appropriate calibration data length when many parameter uncertainties exist. However, as a result of analyzing the parameter sensitivity of the Sacramento model, the variability of the parameter sensitivity decreased as the calibration period increased. The variability of model performance was also related to the variability of parameter sensitivity rather than to the variability of optimal parameter values. Therefore, parameter sensitivity analysis can be used to estimate the calibration data length for various hydrological models, including hydrological models with high uncertainty.