I. Introduction

Hyperspectral imaging sensors provide a thorough description of a scene in a range of wavelengths almost continuous, with a high correlation between adjacent channels [1]. A huge amount of data for each scene is produced, so that problems of transmission, storage, and processing, arise. Generally the whole set of data is redundant to represent the intrinsically embedded information that is sought for [2]. Hence, a representation of the data that allows essential information to be condensed in few basic components is of fundamental importance.