Comparison of soft computing systems for the post-calibration of weather radar

The most usual tools to monitor rainfall events are raingauges and weather radar. Networks of raingauges provide accurate point estimates of rainfall, when appropriately set, but their usual low density restricts considerably the spatial resolution of the gathered information. Such networks, with rain gauges at distinct points, do not reflect the spatial distribution of rainfall. The quality of raingauge observations is also susceptible to some error sources, for example wind effects around the raingauges and poor raingauge reports due to hardware problems. Radar systems offer high spatial and temporal resolution observation which is much more efficient at providing the space-time evolution of a rainfall event in comparison with raingauge networks. However the radar measurements are not free of errors due to a variety of factors including ground clutter, bright bands, anomalous propagation, beam blockages, and attenuation. The effectiveness of weather radar operation is strongly linked to rigorous calibration.; Various methods have been proposed to calibrate radar data. They can be classified into two main categories: deterministic and statistical. The deterministic approach involves the calibration of radar rainfall estimations against raingauge observations. The statistical approach includes multivariate analysis and cokriging. Geostatistical approaches are known as the best methods for radar-raingauge data integration but they are usually inefficient in real time, especially when dealing with the sampling rates of one hour or less necessary for urban and small watershed applications. Such methods also rely on a strong human expertise which can lead to user-dependent results.; The objectives of this research are to introduce and to investigate the feasibility of soft computing systems for the post-calibration of weather radar in comparison with the best existing method based on geostatistics. In this work, the soft computing systems include artificial neural networks and Adaptive Neuro-Fuzzy Inference System (ANFIS) and the geostatistical approach includes residual kriging. The residual kriging calibration results are satisfying however this method is based on stationary hypotheses and requires variogram modeling, making it difficult in an operational context. This method has the advantage of providing a mean squared errors map based on variogram modeling for the estimations. For the artificial neural network, thirteen variants of the multilayer feedforward networks and two variants of radial basis functions are tested in this work. The neural calibration results showed that the Levenberg-Marquardt algorithm using Bayesian regularization is robust and reliable for radar-raingauge data integration. The ANFIS offers the precision and learning capability of artificial neural networks combined with the advantages of fuzzy logic. This method based on the Jackknife approach allows the use of all the available data for training and checking the neuro-fuzzy inference system, and provides a degree of reliability of the post-calibration. The training and the interpolation results of proposed methods can be obtained within just a few seconds using an ordinary personal computer, which is incomparably faster than geostatistical approaches. The proposed algorithms would be very efficient for real time post-calibration.