Investigations in radar rainfall estimation using neural networks

Rainfall on the ground is dependent on the four dimensional distribution of precipitation aloft. In principle one can obtain a functional relation between the rain rate on the ground and the four-dimensional radar observations aloft. However it is difficult to express this in a useful form. Neural networks provide a mechanism to solve this complex problem. Using ground measurements of rain rate as the target output neural networks have been developed in the past that use the radar measurements as input and produce rainfall rates on the ground. Several topics related to neural network based radar rainfall estimates are addressed in this paper. This paper investigates the input vector types and sizes that are useful in a radar rainfall estimation context. Similarly, the neural network is trained with an initial data set, but updated adaptively. Various updating mechanisms are investigated with respect to accuracy of rainfall estimation. Two years of data from the Weather Surveillance Radar-1988 Doppler (WSR-88D) radar and a network of gages from Melbourne, Florida are used to evaluate the topics listed here.