Comparison of four regionalisation methods for a distributed hydrological model

This study presents a grid-based modification of the HBV model concept and four regionalisation approaches using widely available catchment characteristics in the meso-scale Neckar catchment. The HBV model was adapted to allow for the simulation of catchment runoff and daily groundwater recharge in a high spatial discretisation. The resulting large number of model parameters requires the use of a regionalisation method which also ensures consistent parameter estimation. Therefore, in the first approach, functional relationships between catchment characteristics and model parameters have been defined a priori. These established relationships were used to calibrate the model by modifying the parameters of the transfer functions instead of the model parameters themselves. The results are compared to relationships derived from simultaneously calibrated model parameters constrained to form a function of catchment characteristics by a modification of the Lipschitz condition, a monotony condition and a combination of both constraints. Through this reduction of the available parameter space for optimisation, the problem of equifinality is avoided which often results in weak regression relationships between model parameters and catchment characteristics. The methodology is demonstrated using six subcatchments of the Neckar basin to set up the relationships and 51 other subcatchments to evaluate its performance. All four methods were able to produce reasonable parameter sets for most of the regionalisation catchments. As expected, all four methods failed to reproduce the observed discharge in karstic areas and in heavily modified or regulated river basins, which indicates their sensitivity to catchment characteristics. The modified Lipschitz condition produced the most efficient simulations of observed discharges in the regionalisation at the cost of some inconsistencies in the physical interpretation of the resulting relationships. The monotony condition preserved the assumed trends in the functions between cell properties and model parameters but produced sharp jumps which are not considered plausible. The combination of both methods seems to be the most promising because it produced equally good regionalisation results with much more consistent regression relationships. The approach can reproduce derived trends and the resulting relationships match our understanding of how the underlying processes are represented in the model.