Selection of an Artificial Neural Network Model for the Post-calibration of Weather Radar Rainfall Estimation

Abstract

A statistical approach, based on artificial neural networks, is pro posed for the post-calibration of weather radar rainfall estimation. Tested artificial neural networks include multilayer feedforward networks and radial basis functions. The multilayer feedforward training algorithms consisted of four variants of the gradient descent method, four variants of the conju gate gradient method, Quasi-Newton, One Step Secant, Resilient backprop agation, Levenberg-Marquardt method and Levenberg-Marquardt method using Bayesian regularization. The radial basis networks were the radial basis functions and the generalized regression networks. In general, results showed that the Levenberg-Marquardt algorithm using Bayesian regulariza tion can be introduced as a robust and reliable algorithm for post-calibration of weather radar rainfall estimation. This method benefits from the conver gence speed of the Levenberg-Marquardt algorithm and from the over fitting control of Bayes’ theorem. All the other multilayer feedforward training al gorithms result in failure since they often lead to over fitting or converged to a local minimum, which prevents them from generalizing the data. Radial basis networks are also problematic since they are very sensitive when used with sparse data.