Impact of complexity of radar rainfall uncertainty model on flow simulation

A large number of radar rainfall uncertainty (RRU) models have been proposed due to many error sources in weather radar measurements. It is recognized that these models should be integrated into overall uncertainty analysis schemes with other kinds of model uncertainties such as model parameter uncertainty when the radar rainfall is applied in hydrological modeling. We expect that the RRU model can be expressed in a mathematically extensible and simple format. However, the complexity of the RRU has been growing as more and more factors are considered such as spatio-temporal dependence and non-Gaussian distribution. This study analyzes how the RRU propagates through a hydrological model (the Xinanjiang model) and investigates which features of the RRU model have significant impacts on flow simulation. A RRU model named Multivariate Distributed Ensemble Generator (MDEG) is implemented in the Brue catchment in England under different model complexities. The generated ensemble rainfall values by MDEG are then input into the Xinanjiang model to produce uncertainty bands of ensemble flows. Comparison of five important indicators that describe the characteristics of uncertainty bands shows that the ensemble flows generated by MDEG with non-Gaussian marginal and joint distributions are close to the ones with Gaussian distributions. In addition, the dispersion of the uncertainty bands increases dramatically with the growth of the MDEG model complexity. It is concluded that the Gaussian marginal distribution and spatio-temporal dependence using Gaussian copula is considered to be the preferred configuration of the MDEG model for hydrological model uncertainty analysis. Further studies should be carried out in a variety of catchments under different climate conditions and geographical locations to check if the conclusion is valid beyond the Brue catchment under the British climate.