Calibration of global hydrological models based on streamflow characteristics

Streamflow (Q) estimation in ungauged catchments is perhaps the most fundamental challenge faced by hydrologists. A promising approach to calibrate hydrological models in ungauged catchments is to use estimated Q characteristics to identify acceptable model parameter sets. This study is the first to test this approach at the global scale using a large, heterogeneous catchment set. First, observed Q data from 3320 catchments were used to establish neural-network models to derive global maps of five selected Q characteristics, each describing a different aspect of the hydrograph. All Q characteristics could be estimated satisfactorily, with mean training R values ranging from 0.62 to 0.86. Next, a conceptual rainfall-runoff model (HBVLight) was calibrated in a lumped fashion for 200 independent catchments for 2003–2007 using values of the Q characteristics derived from the newly produced maps. A substantial improvement in the simulated Q characteristics was noted, which, in turn, led to improvements in most of the traditional Q performance measures as computed from simulated and observed Q time series. The comparison between the newly produced maps of the respective Q characteristics and estimates derived from two macro-scale hydrological models (Noah and PCRGLOBWB) suggests that the maps can be employed to diagnose the runoff parameterization of the models. The methodology further offers unique possibilities for the diagnosis and/or calibration of future macro-scale hydrological models with high spatial resolution (down to 1 km).