Wavelet based feature reduction method for effective classification of hyperspectral data

As the number of remote sensing applications grows, more intensive is the demand in computationally effective and accurate procedures for spectral analyses. It is known that the computation cost increases with the number of features used for classification. For the maximum likelihood (ML) classifier, the increase is quadratic. Thus, the dimensionality of the initial data is to be reduced prior to analyzing data. In this paper, we propose a new feature extraction and selection method that combines effectiveness of wavelet multiresolution decomposition and the sequential forward floating feature selection (SFFS) technique. The method is used in a combination with the ML classifier. Experimental results show that the developed framework allows for accurate discriminating among similar land cover classes using only few features selected from the initial set of data.