Flood estimation in ungauged catchments: application of artificial intelligence based methods for Eastern Australia

Abstract

Regional flood frequency analysis (RFFA) is used to estimate design floods in ungauged and data poor gauged catchments, which involves the transfer of flood characteristics from gauged to ungauged catchments. In Australia, RFFA methods have focused on the application of empirical methods based on linear forms of traditional models such as the probabilistic rational method, the index flood method and the quantile regression technique (QRT). In contrast to these traditional linear-models, non-linear methods such as artificial neural networks (ANNs) and gene expression programming (GEP) can be applied to RFFA problems. The particular advantage of these techniques is that they do not impose a model structure on the data, and they can better deal with non-linearity of the input and output relationship in regional flood modelling. These non-linear techniques have been applied successfully in a wide range of hydrological problems; however, there have been only limited applications of these techniques in RFFA problems, particularly in Australia. This paper focuses on the development and testing of the ANNs and GEP based RFFA models for eastern parts of Australia. This involves relating flood quantiles to catchment characteristics so that the developed prediction models can be used to estimate design floods in ungauged site. A data set comprising of 452 stations from eastern Australia was used to develop the new RFFA models. An independent testing shows that the non-linear methods are quite successful in RFFA and can be used as an alternative method to the more traditional approaches currently used in eastern Australia. The results based on ANN and GEP-based RFFA techniques have been found to outperform the ordinary least squares based QRT (linear technique).