Bayesian GLS for Regionalization of Flood Characteristics in Korea

For ungauged and gauged sites in Korea, the estimation of flood quantiles is a difficult but important problem. The problem is particularly difficult because almost all flood records are short (< 35 years). The index flood method is commonly used because of the short record lengths, but its performance is compromised by the great heterogeneity among Korean basins. Therefore, a regional flood frequency analysis that addresses differences in the coefficients of variation and skewness using physiographic characteristics of the basin is an attractive approach for constructing the best possible quantile estimators at gauged sites, as well as providing estimates of flood characteristics at ungauged locations. This study presents a Bayesian Generalized Least Square (B-GLS) regression analysis of regional flood frequency data from Korea. The GLS regression framework reflects both the precision of available at-site estimators of flood characteristics and the accuracy of regional models of those statistics. As a result it provides more accurate estimators of model parameters than does ordinary and weighted least square regression analyses, and a nearly unbiased estimator of the model error variance and the precision of estimated parameters. Here B-GLS analyses relate descriptions of scale (the L-CV) and of shape (the L-CS) of the distribution of annual floods to physiographic watershed characteristics, which could include basin index, drainage area, and main channel slope.