Sampling variability in flood frequency analysis: How important is it?

Abstract

Flood quantile estimation using available streamflow records, known as at-site flood frequency analysis (FFA), are widely used in hydrology. The estimated flood quantiles by at-site FFA are used in the planning and design of many water resources management tasks. However, FFA estimates often suffer from high sampling variability, in particular when length of streamflow record is relatively short. This aspect of FFA has not been fully examined for Australian catchments. As the hydrology in Australia suffers from a very high degree of variability, it is likely that the sampling variability in FFA is also very high. This paper presents results from a case study based on three different gauged stations located in New South Wales, Queensland and Victoria using the FLIKE (an extreme value analysis package) software. These stations represent different hydrological regimes (e.g. Victoria is dominated by winter rainfall and Queensland is by summer rainfall). Two widely used probability distribution functions, Generalized Extreme Value (GEV) and Log Pearson type 3 (LP3) distributions are adopted in this case study. We have used updated flood data which have been prepared for Australian Rainfall and Runoff Project 5. The selected streamflow data length ranges from 58 to 102 years. The annual maximum flood data at each of these stations have been sub-divided into three sub-sets: full data set, 50% split and 25% split, which enables to carry out this test with sample sizes in the range of 14 years to 51 years. The study shows that for all the three stations, at-site flood quantile estimates are more affected by the sampling variability in the case of the LP3 (Bayesian) distribution than the GEV (L moments) distribution. Based on the results of this empirical study, it has been found that for 50, 40, 30, 20 and 15 years of annual maximum flood data lengths, the sampling variability estimates are in between -41% to 326% (for LP3 distribution) and -42% to 39% (for GEV distribution) relative to the full data length. The findings of this study have crucial implications in the field of FFA as at-site FFA estimates are generally taken 'accurate' in decision making. Furthermore, in assessing the performances of the regional flood frequency estimation models and calibration of runoff routing model, at-site flood frequency analysis estimates based on about 25 years of data are considered 'robust' and 'accurate', which seems to be not the case. This exercise is being conducted to a greater number of stations by applying boot-strap and Monte Carlo simulation techniques, which will enable to generalize the findings of this study.