Characterization of hydrological model calibration using information theory

Watershed model calibration tunes model parameters to information content found in inputs to fit observations. This tuning has mostly been done using different goodness of fit performance measures (e.g. NSCE, percent bias, etc.). However, such tuning is limited in quantifying and providing insights on how information flows among the various input and output variables and thus whether observed processes are adequately represented within the model. By considering output from uncalibrated and calibrated parameterizations of the National Hydrologic Model run with the Precipitation Runoff Modeling System (NHM-PRMS), this study demonstrated the utility of information theory (IT) in characterizing information flow during model calibration at the H.J. Andrews experimental watershed in Oregon. The results suggest that the calibration process overfits the observed streamflow by poorly extracting the information content of input precipitation and without utilizing the information contained in minimum air temperature beyond the uncalibrated model. As such, the IT-based characterization helps to diagnose the implications of model overfitting, the value of uncalibrated model, and model agility.