An adaptive neural network scheme for precipitation estimation from radar observations

Recent research has shown that neural network techniques can be used successfully for ground rainfall estimation from radars. Application of neural network involves training the network based on past/present data. A neural network may have to be changed with season for best performance. However retraining the network can be a tedious task. In this paper the authors have developed a dynamic neural network which can be changed adaptively with every rainfall regime. A dynamic neural network whose parameters can be adapted in an adaptive manner based on the most recent information is a good compromise solution to the dilemma of accuracy and generalization. A scheme of dynamically updating the structure and parameters of the neural network which enables the network to handle the non-stationary relationship between radar measurements and precipitation estimation with change of season, location and other environment conditions, is developed. The advantages of such a network are shown using data analysis. Data collected by a NEXRAD radar and a network of raingages over Florida is applied to this network to demonstrate the advantage of adaptive neural network for rainfall estimation