Abstract
We compare three approaches to estimate the distribution of extreme rainfall at ungauged sites. Two approaches rely on the univariate generalized extreme value distribution (GEV). SIGEV interpolates linearly the GEV parameters estimated locally. RFA is a regional method which builds circular homogeneous neighbourhood around each site in order to increase the sample size. The observations in the neighbourhood, properly normalized, are assumed to follow the same GEV distribution. Then the normalizing factor (called the index value) has to be interpolated to ungauged sites. The third method is the stochastic hourly rainfall generator called SHYPRE. By characterizing precisely rainfall events, SHYPRE is able to simulate long rainfall series with statistics similar to the observed series. The distribution of extreme rainfall is estimated empirically from the simulated series. The three approaches are evaluated and compared on datasets from over 1000 rain gauges in the South of France. The evaluation framework that we follow is based on the computation of high-level quantiles and aims at assessing the goodness-of-fit of the three approaches and their sensitivity to the training data. Our conclusions are threefold : SIGEV, as implemented, should be avoided because of its lack of robustness, RFA and SHYPRE despite the fact that they are based on very different hypotheses on rainfall provide comparable performance and finally, the main challenge regarding the estimation at ungauged sites concerns the spatial interpolation of the parameters, whatever the approach taken.