In this work, a new strategy for the analysis of hyperspectral image
data is described and assessed. Firstly, the image is segmented into
areas based on a spatial homogeneity criterion of pixel spectra.
Then, a reduced data set (RDS) is produced by applying the projection pursuit (PP) algorithm to each of the segments in which the original hyperspectral image has been partitioned. Few significant spectral pixels are extracted from each segment. This operation allows the size of the data set to be dramatically
reduced; nevertheless, most of the spectral information relative to the whole image is retained by RDS. In fact, RDS constitutes a good approximation of the most representative elements that would be found for the whole image, as the spectral features of RDS are very similar to the features of the original hyperspectral data. Therefore, the elements of a basis, either orthogonal or nonorthogonal, that best represents RDS, are searched for. Algorithms that can be used for this task are principal component analysis (PCA), independent component analysis (ICA), PP, or matching pursuit (MP). Once the basis has been calculated from RDS, the whole hyperspectral data set is decomposed on such a basis to yield a sequence of components, or features, whose (statistical) significance decreases with the index. Hence, minor components may be discarded without compromising the results of application tasks.
Experiments carried out on AVIRIS data, whose ground truth was available, show that PCA based on RDS, even if suboptimal in the MMSE sense with respect to standard PCA, increases the separability of thematic classes, which is favored when pixel vectors in the transformed domain are homogeneously spread around their class centers.