Inventaires de crues extrêmes sur des sites non jaugés et analyse statistique régionale des débits: réflexions méthodologiques et évaluation des performances

Abstract

This paper presents additional developments about a flood frequency approach proposed by Gaume et al. [2010], aiming at incorporating available information on extreme floods at ungauged sites in a regional flood frequency analyse. Extreme floods, indeed, may represent a very large additional record length if compared to gauged series, and therefore may highly improve the accuracy of flood quantiles estimates. However, the proposed approach requires to calibrate an index flood relationship (here an exponent of catchment surface), which complexifies the regional statistical model. The performances and robustness of this approach are tested here and compared to the conventional regionalisation approach of Hosking & Wallis [1997]. The comparison is based both on simulations and on a case study. The inference procedure used is based on a GEV distribution associated with a specific likelihood formulation and a Bayesian MCMC inference procedure for parameters estimation. The parameters of the index flood relationship and the GEV distribution are all calibrated in the same time using the MCMC algorithm, and therefore the uncertainty associated with the index flood relationship calibration is taken into account. The inference results obtained without incorporing extreme events are first compared, with focus on the effects of the index flood relationship estimation and on the presence of heterogeneities around this relationship. This first comparison is based on simulations and illustrates some limitations of both of the methods: sensitivity to the length of local series in the case of the Hosking & Wallis approach, presence of bias for small catchment surfaces and relative sensitivity to heterogeneities in the case of the proposed approach. As far as the available information is limited to gauged series and if heterogeneity remains limited, the two approaches finally show similar performances. The two approaches are then applied to a case study: the Ardèche region in south east of France (168 gauged records at 5 sites and 18 ungauged extremes). This application confirms the very positive effect of the information on extreme floods, which enables to outperform the results based only on gauged series. © Société Hydrotechnique de France, 2013.