Extraction of spectral channels from hyperspectral images for classification purposes

This paper proposes a procedure to extract spectral channels of variable bandwidths and spectral positions from the hyperspectral image in such a way as to optimize the accuracy for a specific classification problem. In particular, each spectral channel ("s-band") is obtained by averaging a group of contiguous channels of the hyperspectral image ("h-bands"). Therefore, if one wants to define m s-bands, the problem can be formulated as the optimization of the related m starting and m ending h-bands. Toward this end, we propose to adopt, as an optimization criterion, an interclass distance computed on a training set and to generate a sequence of possible solutions by one of three possible search strategies. As the proposed formalization of the problem makes it analogous to a feature-selection problem, the proposed three strategies have been derived by modifying three feature-selection strategies, namely: 1) the "sequential forward selection", 2) the "steepest ascent," and 3) the "fast constrained search". Experimental results on a well-known hyperspectral data set confirm the effectiveness of the approach, which yields better results than other widely used methods. The importance of this kind of procedure lies in feature reduction for hyperspectral image classification or in the case-based design of the spectral bands of a programmable sensor. It represents a special case of feature extraction that is expected to be more powerful than feature selection. The kind of transformation used allows the interpretability of the new features (i.e., the spectral bands) to be saved.