On the use and misuse of semi-distributed rainfall-runoff models

Recent advances in hydrological modelling have led to a variety of complex, distributed or semi-distributed schemes, aiming to describe the heterogeneity of physical processes across a river basin. These are useful for operational purposes, such as design of large hydraulic structures, sustainable management of water resources and flood forecasting. However, due to the large number of parameters involved and the need for extended measurements, a robust calibration, which ensures a satisfactory predictive capacity as well as a physical interpretation of parameters, is a very difficult task. Hence, the applicability of such models in real-world studies, employed by practitioners with moderate hydrological knowledge, is at least questionable. The paper aims to reveal some critical issues, regarding the entire procedure of selecting, configuring and fitting a hydrological model. These are discussed on the basis of four classification criteria: the expertise level of the user, the representation of processes, the parameterization concept and the calibration strategy. An inexperienced user focuses on just finding a good fitting between model outputs and observations, usually by activating more parameters than are supported by the data. In contrast, an expert hydrologist wishes to explain the entire spectrum of model results, giving emphasis on the reasonable representation of the processes and the consistency of the all output variables, even those not controlled by the calibration (e.g. real evapotranspiration, soil moisture and groundwater storage fluctuation, etc.). In terms of the processes representation, modelling approaches that are devised for uniform, undisturbed basins are misused if applied on complex systems, with multiple human interventions. The next criterion refers to the parameterization procedure. Some approaches assign parameter values on the basis of the schemati zation, i.e. the spatial discretization of the system under study (e.g. the sub-basins), thus leading to schemes with too many degrees of freedom, suffering from the well-known "curse of dimensionality". On the other hand, more intelligent models assume different levels of parameterization and schematization, employing the concept of a hydrological response unit. Thus, they significantly reduce the number of control parameters, also ensuring consistency with the physical characteristics of the system under study. Finally, one may classify the calibration strategies from manual, one-criterion fitting to sophisticated automatic optimization methods, using evolutionary algorithms and multiple fitting criteria, both statistical (based on measurements) and empirical (based on the hydrological experience). The above spectrum of modelling options is explored by selecting representative cases which reveal problems of everyday hydrological practice. The test area is the Boeoticos Kephisos basin, Greece, where a conjunctive simulation model is employed to describe the surface and groundwater hydrological processes as well as the water management practices.