Regional flood frequency model for Kigoma and Rukwa regions, western Tanzania.

A regional flood frequency model - useable at both gauged and ungauged sites in Kigoma and Rukwa regions - western Tanzania - is developed. This is done using the methods of hydrological (stochastic) modelling, which entail fitting a statistical model to a regionally pooled sample data. Twenty - one (21) representative catchments from both Kigoma and Rukwa regions were selected to represent river catchments in those regions and their annual maximum flood (AMF) flows pooled together in the analysis. This study examined the most appropriate of three widely used statistical models capable of fitting the standardized regional data. The probability weighted moment method of model parameter estimation was used. Results show that the extreme value type 1, general extreme value and Iog-logistic distributions satisfactorily fitted the data, as they all passed the Chi-square goodness of fit test. However EV1 is the best for the regional flood frequency model {RFFM} with respect to effectiveness, efficiency and ease of computation, To be able to use the RFFM for ungauged sites, a model for predicting index floods from measures of physical catchment characteristics is developed, using logarithmic regression techniques it was found that of the 12 combinations of the measures - catchment area, annual rainfall maximum rainfall for one month, maximum rainfall for two consecutive months and maximum rainfall for three consecutive months - catchment area and annual rainfall best modelled the relationship. As a test of homogenity the outlier influence was detected. It is concluded from the analysis of outlier that the log-logistic distribution is more robust than the other candicate distributions in dealing with outlier at the upper end of flood frequency series. Risk from using short record was calculated. It was found that the risk ratio decreases rapidly with the addition of a few years of record if N is small, while the same amount of increase on relatively large samples does not improve the result significantly.