Model Performance Sensitivity to Objective Function during Automated Calibrations

Abstract

Previous studies have reported limitations of the efficiency criteria commonly used in hydrology to describe goodness of model simulations. This study examined sensitivity of model performance to the objective function used during automated calibrations. Nine widely used efficiency criteria were evaluated for their effectiveness as objective function, and goodness of the model predictions were examined using 13 criteria. Two cases (Case I: Using observed streamflow data and Case II: Using simulated streamflow) were considered to accomplish objectives of the study using a widely used watershed model (SWAT) and good-quality field data from a well-monitored experimental watershed. Major findings of the study include (1) automated calibration results are sensitive to the objective function group—group that work based on minimization of the absolute deviations (Group I), group that work based on minimization of square of the residuals (Group II), and groups that use log of the observed and simulated streamflow values (Group III)—but not to objective functions within the group; (2) efficiency criteria that belong to Group I were the most effective when used as objective function for accurate simulation of both low flows and high flows; (3) Group I and Group II objective functions complement each other’s performance; (4) with regard to the capability to describe goodness of model simulations, efficiency criteria that belong to Group I showed superior robustness; (5) for the study watershed, use of the long-term interannual calendar day mean as baseline model did not improve capability of an efficiency criterion to describe model performance; and (6) even for ideal conditions where uncertainty in input data and model structure are fully accounted for, identifying the so-called global parameters values through calibration could be daunting as parameter values that were significantly divergent from predetermined values produced model simulations that can be considered near perfect even when judged using multiple efficiency criteria.