frequency analysis

Flood observations in hydrological data sets contain frequently outliers, and this causes problems for water resource researchers and planners if not addressed correctly. This study analyses how outliers affect the identification of regional probability distributions using L-moment methods. The main objective of the study is to assess the effect(s) of discordancy detection measures on regional flood probability types and the accuracy of the estimates based on the regional analysis. The classical and robust discordancy measures for discordant site identification are used to determine regional probability distributions in order to identify the effects of discordant sites on the regional probability distribution in a region of the Menderes River Basins in Turkey. The other objective is to show whether a probability model type and flood estimation based on the model is reliable if discordancy sites in the region are not detected. In the study, the homogeneity of the basin was tested using the L-moments based on the heterogeneity for two discordancy measures, assessed by carrying out 500 simulations using the four parameter Kappa distribution. Based on these tests, two sub-regions are defined, the Upper-Menderes and Lower-Menderes sub-regions, that have different numbers of sites for both discordancy measures. According to the L-moments goodness of statistic criteria, the generalized extreme value distribution was determined as the best-fit distribution for the Upper-Menderes and Lower-Menderes sub-regions based on the classical discordancy measure. The generalized extreme value distribution was also found to be the best-fit distribution for the Upper-Menderes sub-region for the robust discordancy measure, while the Pearson Type 3 distribution was the best for the Lower-Menderes sub-region based on the robust measure. To appraise the results for the sub-regions, the relative root mean square error and relative bias were employed. The results show that the homogeneous region determined from the robust discordancy measure is more accurate than the region identified using the classical robust measure. This means that the classical robust detection measure of flood frequency analysis needs to be improved. (C) 2009 Elsevier B.V. All rights reserved