Design of spectral channels for hyperspectral image classification

Purpose of this paper is to propose a procedure to extract, from a hyperspectral image, spectral channels of variable bandwidths and spectral positions 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"). If one wants to define it s-bands, the problem can be therefore formulated as the optimization of it starting and it ending h-bands. To this end, we propose to adopt as optimization criterion an interclass distance computed on a training set and to generate a sequence of possible solutions with one of three possible search strategies. As the proposed formalization of the problem makes it analogous to a feature selection problem, the three proposed strategies have been derived by modifying three feature selection strategies: the Sequential Forward Selection (SFS), the Steepest Ascent and the Fast Constrained Search. Experimental results with a well-known hyperspectral data set confirm the effectiveness of the approach, which allows better results than those provided by the SFS method for feature selection. A preliminary comparison suggests that the accuracy is very similar to that obtained by the DBFE feature transformation method. The interest of this kind of procedure can be for a case-based design of the spectral bands of a programmable sensor or for the reduction of the number of bands derived from a hyperspectral image. It represents a special case of feature transformation 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.