'Outlier' catchments: what can we learn from them in terms of prediction uncertainty in rainfall-runoff modelling?
Résumé
What exactly are the catchments that we usually exclude from our data sets before submitting a paper to a conference, on the grounds that our models fail to represent their behaviour? What can be the consequences of the commonly-used (but rarely discussed) data set cleansing practise: does it help us improve our models? Does it contribute to making our hydrological simulation less uncertain? Or does it just give us a false sense of confidence in our capacity to represent catchment hydrological behaviour? This paper focuses specifically on the "outlier" catchments found in a large set of 1045 French catchments. This large catchment set allows statistical quantification of the likely sources of model failure ; it shows regional clustering (linked with the geology), the surprising effect of catchment area (the largest basins get the best performances) and last, that noise in input data is in no way sufficient to explain the difficulties of five rainfall-runoff models in representing catchment behaviour.