MULTIRESOLUTION ANALYSIS FOR SEPARATING CLOSELY SPACED FREQUENCIES WITH AN APPLICATION TO INDIAN MONSOON RAINFALL DATA

Abstract

In this paper we make use of the multiresolution properties of discrete wavelets, including their ability to remove interference, to reveal closely spaced spectral peaks. We propose a procedure which we first verify on two test signals, and then apply it to the time series of homogeneous Indian monsoon rainfall annual data. We show that, compared to empirical mode decomposition, discrete wavelet analysis is more effective in identifying closely spaced frequencies if used in combination with classical power spectral analysis of wavelet-based partially reconstructed time series. An effective criterion based on better localization of specific frequency components and accurate estimation of their amplitudes is used to select an appropriate wavelet. It is shown here that the discrete Meyer wavelet has the best frequency properties among the wavelet families considered (Haar, Daubechies, Coiflet and Symlet). In rainfall data, the present analysis reveals two additional spectral peaks besides the fifteen found by classical spectral analysis. Moreover, these two new peaks have been found to be statistically significant, although a detailed discussion of testing for significance is being presented elsewhere.