Precipitation estimation and forecasting using radar and rain gage measurements with artificial neural networks

Quantitative precipitation estimation and forecasting continue to be critical components of the weather research programs. The objective of this dissertation is twofold: First, to propose a method that fuses rainfall measurements from rain gages and radar. Second, to design a technique that produces real-time rainfall forecasts for the next hour. Cokriging is perhaps the most widely used method to fuse measurements from two sensors, for example, radar and rain gages. Here an alternative fusion methodology, based on recent developments in Artificial Neural Networks (ANNs) is presented. ANNs are nonlinear estimators and thus have a distinct advantage over traditional statistical methods. Intercomparison of rainfall estimation, using cokriging and ANN methods, suggests that ANNs provide a more attractive and robust fusion from radar and rain gages for several storms from Oklahoma.; It is shown that simply nowcasting the fused estimates gives better forecasts than the traditional nowcasting with radar data. Moreover, the rainfall field at the next hour is predicted with a methodology that is based on radial-basis ANNs. The advantage of this method is that it provides a framework for the automated segmentation of the rainfall field in rainfall clusters that have their own advection vectors. Each cluster is shifted individually for the prediction step. Thus, the method accounts for nonhomogeneous advection conditions. The results show that this method has the capability to generate improved predictions compared to nowcasting. It appears that its full strength will be realized if a data set with a temporal resolution finer than hourly is used. In summary, an integrated ANN approach has been produced that estimates rainfall from two sensors and produces a forecast.; In the appendix I also include a study on the nature of long-range rainfall and streamflow correlations, using a method called Detrended Fluctuation Analysis. The findings show the existence of power-law correlations in both variables. Moreover, it is shown that what controls the correlation structure is not the actual rainfall values, but the pattern of alternating wet and dry spells. The employed method also highlights the dampening effect of the soil in the transformation of rainfall to streamflow.