Calibration of hydrological model GR2M using Bayesian uncertainty analysis

An outstanding issue of hydrological modeling is the adequate treatment of uncertainties in model calibration and prediction. The current paradigm is that the major sources of uncertainties, namely input, output and model uncertainty should be accounted for directly, instead of assuming they can be safely lumped into the output uncertainties. In this paper, Bayesian analysis is used to calibrate the conceptual hydrologic monthly model GR2M taking into account input, output, structural and initial state uncertainty through error models and priors. Calibration is performed under different error assumptions to study the influence of the initial state uncertainty, the consequences of large input errors, the impact of error assumptions on calibrated parameter posterior distributions and the definition of error models. It is shown how such an analysis can be used to separate, a posteriori, the different sources of errors, and in particular, to identify structural errors from data errors.