Feature Extraction Via Multiresolution MODWT Analysis in a Rainfall Forecast System

Abstract

During 30 years, expert meteorologists have been sampling meteorological measurements directly related to the rainfall event, in order to improve the current forecast procedures. This study performs the Feature Extraction and Feature Selection processes to extract the relevant information in the rainfall event. The Feature Extraction has been performed with a Multiresolution Analysis applying the Maxima OverlapWavelet Transform. The selection of the wavelet decomposition, was obtained applying a Sequential Feature Selection algorithm based on General Regression Neural Networks. In this paper, it is also presented a novel architecture to perform short and medium term weather forecasts based on Neural Networks and time series estimation filters. The preliminary results obtained, present this architecture as a feasible alternative to the current forecast procedures performed by super computer simulation centers.