Testing similarity indices to reduce predictive uncertainty in ungauged basins
Résumé
In ungauged watersheds, the common approach for estimating the parameter values of lumped rainfall-runoff models consists of a two-step approach First, relationships between watershed characteristics and parameter values are established on gauged sites, and second, these relationships are used to estimate parameter values on ungauged sites However, several studies suggested that there are strong limitations to this approach and that consideration for similarity and/or proximity offered a better outlook. We propose here an original approach based both on similarity considerations and multi-model methodology. First, gauged watersheds are clustered into 27 classes depending on the values of three characteristics (either physical or hydro-climatic). Then, for each ungauged watershed, we used the calibrated parameters of similar gauged watershed, i.e. from the same class. Then, a combination of the simulations obtained with the different sets of parameters was performed. The methodology is based on the GR4J rainfall-runoff applied on more than 1000 basins located in France Results show that the physical similarity approach performs slightly better than the regression-based approach Refinements of these tow approaches, such as regional calibration of regressions or multi-model considerations for regionalization based on physical similarity, do not yield significant improvements.