Towards reducing model error in flow predictions in ungauged basins via a Bayesian approach

Flow predictions in ungauged basins (PUBs) remain an elusive challenge in hydrological sciences and engineering, even with the advances achieved during the "PUB decade" (Hrachowitz et al. 2013). Meeting this challenge is made difficult by the uncertainty in the "regionalization" model used to transpose hydrological data from gauged to ungauged basins (e.g. Wagener and Wheater 2006), and by the uncertainty in the hydrological model used to predict streamflow in the ungauged basin (Clark et al. 2011). Traditionally, regionalization proceeded by transferring the parameters of a hydrological model from gauged basins to the ungauged basin. A more recent approach to regionalization is to extrapolate the hydrological characteristics (flow indices) of a catchment rather than its fitted hydrological model parameters (see Wagener and Montanari 2011). Many advances have been made in the last decade in flow indices selection (e.g. using principal components analysis; Olden and Poff 2003) and regionalization (e.g. regression via Random Forest instead of linear regression; Snelder et al. 2009). Once regionalized flow indices are available, conditioning the hydrological model parameters on regionalized flow indices via a Bayesian approach has proven to be advantageous to quantify and reduce the uncertainty in the hydrological model streamflow predictions (e.g. Bulygina et al. 2009; Blöschl et al. 2013). The approach rests on the assumptions that both regionalization model and hydrological model are adequate for a selected ungauged catchment; and hence requires devising tests to evaluate the assumptions (Wagener and Montanary 2011).