Unsupervised band selection for multispectral images using information theory

In this paper, the implication of the relations of information in the case of multispectral images is analyzed. Higher-order mutual information can adopt positive or negative values depending of the correlation among ensembles. Therefore, the existence of negative values reflects higher-order correlations in the conditional information. On the other hand, the extraction of optimal subsets of spectral images is proposed as a maximization of the conditional entropies at same time that the dependent information among images is minimized.