Operational Feasibility of Neural-Network-Based Radar Rainfall Estimation

Abstract

An operational radar rainfall estimation system based on the adaptive radial basis function (RBF) neural network is developed. During the process of training and cross validation, the rainfall estimation was computed only at the gauge locations. Once the training is done, the radar rainfall estimation based on neural networks is applied to the full coverage area of the radar. Such large-scale application of the rainfall estimate poses several questions in the context of operational applications. This letter addresses two of those questions, namely: 1) the feasibility of adaptively updating RBF neural network models on a daily basis and 2) the ability of neural network radar rainfall estimation at high spatial resolution within reasonable and practical time frame for operational applications. Using the datasets collected by WSR-88D radar located in Melbourne, FL, it is demonstrated that radar-based rainfall estimation using an adaptive RBF neural network is feasible. The results show that 73% of overnight updating for the RBF neural network can be completed within 2 h, and the estimation over an area of 100 km/spl times/100 km can be generated within the time frame (a few tens of seconds-150 s), which is much smaller than the average radar volume scan time.