Neural network applications in weather radar systems

Accurate estimation of ground rainfall from radar measurements is an important topics of current interest. Traditionally, radar rainfall estimates were computed from a parametric reflectivity-rainfall (Z-R) relation that was allowed to vary from place to place and season to season. Such an approach has not been very successful because of the extensive variability observed with Z-R relations with rain type and climatic regions. With the advance of multiparameter radar, the rainfall estimation problem is addressed by obtaining better characterization of the precipitation medium using differential reflectivity ({dollar}Zsb{lcub}dr{rcub}{dollar}) and specific differential propagation phase ({dollar}Ksb{lcub}dp{rcub}{dollar}). However only few experiments done in a controlled manner were able to demonstrate significant improvements using multiparameter radar based rainfall estimates. The problems that exist in the comparison of estimates from radar with observations from ground based instruments suggest that both radar and surface observations may depend on the three-dimensional structure of the drop size distribution.; Neural network based techniques using three dimensional radar measurements to obtain precipitation estimates on the ground are introduced in this study. Multilayer perceptron (MLP) networks with recursive least square (RLS) back propagation learning algorithm have been developed to obtain rainfall estimates based on data collected by CP-2 multiparameter radar and NEXRAD (Next Generation Weather Radar) systems. Both the surface radar measurements and vertical profiles are applied to the network as input. And the ground raingage network measurements are used as both the target outputs during the training and the standard results during the performance evaluation process. For data from the multiparameter radar, both the reflectivity ({dollar}Zsb{lcub}h{rcub}){dollar} and the differential reflectivity ({dollar}Zsb{lcub}dr{rcub}){dollar} are used to develop the network to obtain rainfall estimates. Results are compared with existing Z-R and {dollar}Zsb{lcub}h{rcub},{dollar} {dollar}Zsb{lcub}dr{rcub}{dollar}-R relations to show significant improvements. For data from the NEXRAD, a Kohonen self-organization feature mapping (SOFM) network has been developed to identify the presence/absence of precipitation on the ground based on vertical reflectivity profiles. Neural network rainfall estimates are compared with 24-hour raingage accumulations over a whole month. Significant improvements are observed by comparing the performance of neural network rainfall estimation algorithm to other conventional techniques.; Finally, a radial-basis function (RBF) neural network is developed for the snowfall estimation problem. Based on the data collected by the CSU-CHILL multiparameter radar and snowgages located at the Stapleton International Airport (SIA) and Denver International Airport (DIA), the RBF network is trained and tested to estimate ground snowfall depth using vertical reflectivity profiles. Through several different training and testing schemes, the RBF networks show good promise to map the functional relationship between the radar observed vertical profiles and ground snow depth.