Résumé
A French national project has been set up since 2002 in order to study the variation of floods and droughts in France, based on a large set of discharge data series. Numerous approaches can be used to study trends and variations on hydrological data series, e.g. statistical tests, segmentation procedures, Bayesian approaches, and multivariate or regional models. A comparative analysis aimed at defining a general framework for the selection of a limited number of tests, based on the statistical properties of the hydrological data series (sample size, autocorrelation) and the prior knowledge of the type of distribution and the expected change. The first results on a set of about 200 French long data series do not show conclusive proof that climatic change has affected flood and drought regimes. A Bayesian model has finally been developed to incorporate hydrological nonstationarity into frequency analysis, using MCMC algorithms for the weighting of different time dependent models.