Features of regional flood frequency estimation (RFFE) model in Australian rainfall and runoff

![CDATA[Flood estimates at ungauged catchments are generally associated with a high degree of uncertainty. For the upcoming Australian Rainfall and Runoff (ARR) 2015 (4th edition), a new regional frequency estimation (RFFE) model known as ARR RFFE Model 2015 has been developed. This RFFE Model 2015 is based on the concept of regionalization, which is a data driven approach where data from gauged catchments are utilized to make flood quantile estimates at ungauged locations. This paper presents the modelling approach that underpins the RFFE Model 2015. In the RFFE Model 2015, Australia is divided into humid coastal areas and arid/semi-arid areas. In the humid coastal areas, a region of influence approach is adopted to form sub-region by drawing a number of nearby gauged catchments for a location of interest. This in essence attempts to reduce the degree of heterogeneity in forming local region at the location of interest. For estimating flood quantiles, a regional log Pearson Type 3 (LP3) distribution is adopted where the location, scale and shape parameters are estimated based on prediction equations. A Bayesian Generalized Least Squares (GLS) regression approach is adopted to develop prediction equations in the humid coastal areas. The main advantages of GLS regression is that this accounts for inter-station correlation and variation in streamflow data lengths across different sites in a region. Furthermore, GLS regression differentiates between the sampling and model errors and hence provides a more rigorous approach of dealing with uncertainty in regional flood modelling compared with the ordinary least squares regression. For the arid/semi-arid areas, an index flood method is adopted where 10% annual exceedance probability (AEP) flood is taken as the index variable. The prediction equation for the index variable is developed using an ordinary least squares regression approach in the arid/semi-arid areas. The regional growth factors are estimated from the at-site flood frequency analysis where LP3 distribution is fitted to the annual maximum flood series data. The Multiple Grubb-Beck test is used to censor the zero and low annual maximum flood series data points, which is typical in the arid regions. A total of 798 gauged catchments are used from the humid coastal areas and 55 catchments from the arid/semi-arid areas to develop and test the ARR RFFE model. The data from these catchments are prepared adopting a stringent quality control procedure that involved infilling the gaps in the data, checking for outliers, trends and rating curve extrapolation error. In developing the confidence limits for the estimated flood quantiles, a Monte Carlo simulation approach is adopted by assuming that the uncertainty in the first three parameters of the LP3 distribution (i.e. the mean, standard deviation and skewness of the logarithms of the annual maximum flood series) can be specified by a multivariate normal distribution. In the ARR RFFE model, the model coefficients have been embedded in an application software (known as RFFE Model 2015), which enables the user obtaining design flood estimates relatively easily using simple input data such as latitude, longitude and catchment area of the ungauged catchment of interest. The RFFE Model 2015 is applicable to any catchment that has similar attributes and flood producing characteristics as the catchments used in the derivation of the flood estimation equations embedded in the RFFE model. Catchments which do not satisfy this requirement can be divided into three groups: (i) catchments which have been substantially modified from their natural characteristics and for which the RFFE model is not applicable and should thus not be used (ii) catchments for which flood estimates must be expected to have lower accuracy such as arid region catchments; and (iii) ‘atypical catchments’ where additional catchment attributes need to be considered and adjusted for such as catchments with large natural flood plain area.]]