Hyperspectral Images Classification with Typical Sequences associated to the Endmember

摘要

This paper presents a new methodology for hyperspectral image classification based on the definition of typical sets from the Asymptotic Equipartition Property, an important tool in the field of information theory. The Endmembers (EM) are decomposed in orthogonal functions by a discrete wavelet transform and are modeled as a HMM (Hidden Markov Model). Based on this model, for each EM, a Typical Sequence set is established. One spectrum is classified as a member of a specific EM if belongs to its typical set. It is considered the case in which a class in the hyperspectral image can be represented by several subclasses and also the original spectra can be decimated and be used with less bands in the classification processes. The proposed method is tested with a set of AVIRIS data and is compared with the classification performed by Euclidian Distance, Spectral Angle Mapper (SAM) and Spectral Information Divergence (SID). It is shown that the proposed classification can be used with a reduced number of bands and achieves results comparable with other methods using all bands.