Wavelet Based Feature Extraction Techniques of Hyperspectral Data

Hyperspectral data have many applications and are being promoted over multi-spectral data to derive useful information about the earth surface. But this hyperspectral data suffers from dimensionality problem. It is one of the challenging tasks to extract the useful information with no or less loss of information. One such technique to extract the useful information is by using wavelet transformations. In this paper, a series of experiments have been presented to investigate the effectiveness of some wavelet based feature extraction of hyperspectral data. Three types of wavelets have been used which are Haar, Daubechies and Coiflets wavelets and the quality of reduced hyperspectral data has been assessed by determining the accuracy of classification of reduced data using Support Vector Machines classifier. The hyperspectral data has been reduced upto four decomposition levels. Among the wavelets used for feature extraction Daubechies wavelet gives consistently better accuracy than that produced from Coiflets wavelet. Also, 2-level decomposition is capable of preserving more useful information from the hyperspectral data. Furthermore, 2-level decomposition takes less time to extract features from the hyperspectral data than 1-level decomposition.