A comparison of regional flood frequency estimation methods using a resampling method

Most comparisons of flood frequency estimation methods have been based on Monte Carlo simulation. An alternative method which does not require the choice of a parent distribution is to resample from the historical record. The relative performance of 10 flood frequency estimation methods, including four at-site methods, four regional index flood methods, and two variations of the U.S. Water Resources Council regional method, was evaluated using a resampling method. Performance measures were based on two statistics. The first was a flood damage surrogate, which was defined as a power of the excess of the flood peak over a fixed threshold approximately equal to the median annual flood; the second was the estimator of the 100-year flood. For each subsampling of the historical record, the expected flood loss and 100-year flood quantile conditioned on the estimated parameters were computed. From the remainder of the record, an empirical estimate of the mean damage was computed. The performance of the methods is summarized in terms of the mean and variance of the difference between the conditional expectation and the empirical damages, and the variance of the 100-year flood estimates. The results are qualitatively similar to those reported in earlier Monte Carlo studies. The U.S. Water Resources Council method, in particular, produced variability in the damage estimates that was much greater than that for any of the index flood methods tested. Estimates of the 100-year flood were about twice as variable on average as 100-year flood estimates obtained using any of the regional index flood methods. For the Wisconsin and New England flood records tested, the regional Generalized Extreme Value estimators yielded 100-year flood estimates with lower variability at virtually all sites than the U.S. Water Resources Council method.