Parameter optimization and uncertainty analysis for plot-scale continuous modeling of runoff using a formal Bayesian approach

This study reports on the use the recently developed Differential Evolution Adaptative Metropolis algorithm (DREAM) to calibrate in a Bayesian framework the continuous, spatially distributed, process-based and plot-scale runoff model described by . The calibration procedure, relying on 2 years of daily runoff measurements, accounted for heteroscedasticity and autocorrelation. This resulted in a realistic estimation of parameter uncertainty and its impact on model predictions. The calibrated model reproduced the observed hydrograph satisfactorily during calibration. The model validation on an independent 1-year series of measurements showed reasonable values for the Nash–Sutcliffe efficiency criterion. This was equal to 0.76 and 0.78 for, respectively, the upper and lower bounds of the 95% uncertainty interval associated with parameter uncertainty. However, the fact that this interval was always very narrow but often did not bracket the observations during the validation period indicates that, assuming observational errors to be negligible, model structure has to be improved in order to achieve more accurate predictions. Overall the methodology allowed to efficiently discriminate between parameter and total predictive uncertainty and to highlight scope for model improvement.