The use of artificial neural networks (ANNs) in modelling the hydrological processes has become a common approach in the last two decades, among side the traditional methods. In regard to the rainfall-runoff modelling, in both traditional and ANN models the use of ground rainfall measurements is prevalent, which can be challenging in areas with low rain gauging station density, especially in catchments where strong focused rainfall can generate flash-floods. The weather radar technology can prove to be a solution for such areas by providing rain estimates with good time and space resolution. This paper presents a comparison between different ANN setups using as input both ground and radar observations for modelling the rainfall-runoff process for Bahluet catchment, with focus on a flash-flood observed in the catchment.
The rainfall-runoff transformation is a highly complex dynamic process and the development of fast and robust modelling instruments has always been one of the most important topics for hydrology. Over time, a significant number of hydrological models have been developed with a clear trend towards a process-based approach. The downside of these types of models is the significant amount of data required for building the model and for the calibration process: in practice, the collection of all necessary data for such models proves to be a difficult task. In order to cope with this issue, various data-driven modelling techniques have been introduced for hydrological modelling as an alternative to more traditional approaches, on the basis of their capacity of mapping out complex relationships from observation data. Having the capacity to generate meaningful mathematical structures as results, genetic programming (GP) presents a high potential for rainfall-runoff modelling as a data-driven method. Using ground and radar rainfall observation, the aim of this study is to investigate the GP technique capability for modelling the rainfall-runoff process, taking into consideration a flash-flood event.