Testing Flood Frequency Estimation Methods Using a Regional Flood Generation Model

Abstract

A regional flood generation model is proposed, wherein the logarithms of the first two moments of the annual flood distribution at each site depend on the logarithm of the drainage area. This model is parameterized in terms of the regional mean coefficient of variation, , and the site‐to‐site coefficient of variation of the coefficient of variation, C c . Eight flood frequency estimation methods, including two at‐site and six regional estimators, were tested. Generated floods were drawn from three parent distributions: extreme value l, lognormal, and Pearson 3. The regional estimators had lower root mean square errors than the at‐site estimators, even when the region was moderately heterogeneous (high C c ), except when the parent was extreme value 1. However, the regional estimators were not always robust with respect to the parent flood distribution. The flood index method, which was based on regional probability weighted moments, performed quite well when and C c were both low. As and C c , increased, however, its performance degraded markedly, primarily due to dominance of the quantile estimation variance by the variance in the at‐site mean estimator. This result suggests that for high and/or C c , improvements in regional flood estimation will come from improved estimators of the at‐site mean annual flood, rather than the regional (normalized) flood frequency distribution.