Abstract
In the French Mediterranean area where heavy precipitation events canyield devastating consequences, it is essential to obtain reliableestimates of the distribution of extreme precipitation at gauged andungauged locations. Under mild assumptions, extremes defined asexcesses over a high enough threshold can be modeled by thegeneralized Pareto (GP) distribution. The shape parameter of the GPwhich characterizes the behavior of extreme events is notoriouslydifficult to estimate. In regional analysis, the sample variabilityof the shape parameter estimate can be reduced by increasing thesample size. This is achieved by assuming that sites in a so-calledhomogeneous region are identically distributed apart from a scalingfactor and therefore share the same shape parameter. A majordifficulty is the proper definition of homogeneous regions. We buildupon a recently proposed approach, based on the probability weightedmoment (PWM) for the GP distribution, that can be cast into a regionalframework for a single homogeneous region. Our main contribution isto extend its applicability to complex regions by characterizing eachsite with the second PWM of the scaled excesses. We show on syntheticdata that this new characterization is successful at identifying thehomogeneous regions of the generative model and leads to accurate GPparameter estimates. The proposed framework is applied to 332 dailyprecipitation stations in the French Mediterranean area which aresplitted into homogeneous regions with shape parameter estimatesranging from 0 to 0.3. The uncertainty of the estimators is evaluatedwith an easy-to-implement spatial block bootstrap.