Rainfall-runoff (RR) models have a wide range of operational applications. Among these applications, streamflow forecasting probably enjoys the highest demand worldwide. This domain is very demanding, as it requires both quality and robustness. Lumped RR models have given numerous proofs of their robustness, and they can be used with efficient updating algorithms to provide accurate forecasts. On the other hand, distributed models have more possibility to take into account the spatial heterogeneity of both rainfall and soils, and this could contribute to the improvement of streamflow forecasts.

This paper attempts to introduce a new way to help modellers decide of the most appropriate equilibrium between lumped and distributed approaches in RR modelling. We compare lumped and semi-distributed approaches over a large number of watersheds, and test a range of explanatory factors to try to explain the differences in model efficiency in both modes.

To do so, we use a database of 300 French watersheds, for which rainfall, runoff and potential evapotranspiration are available at the daily time step. We construct what we call “chimera watersheds” by associating two actual watersheds similar in size: runoff of the resulting watershed is obtained by addition of sub-watersheds runoff, and rainfall input is a weighted average of sub-watersheds rainfall. We produce a total of 4,500 combinations, for which we compare the performances of the lumped and of various semi-distributed approaches.

We believe that the use of “chimera watersheds”, by providing very contrasted hydrological situations, can be useful to help identify those factors which are relevant to determine the most appropriate level of spatial distribution for a RR model. For this study, of particular interest is the fact that the largest part (3/4th) of the improvement brought by distribution is due to rainfall variability.