Image Processing Reference
Band Selection: Revisited
In last few chapters we have seen four different methodologies for fusion of hyper-
some of the salient characteristics of the input data. In the last chapter, we have
discussed alternately how the fusion technique can be based on the desired charac-
teristics of the output. For most of the users and analysts, the visual quality of the
fused image is of primary importance. The key issue is then how to efficiently capture
the features from the input images and transfer them into saliency of the fused image.
These features are mainly application dependent, and thus, the technique discussed
in Chap. 7 is more of a generic nature. The idea behind the optimization-based fusion
technique described in the last chapter is to develop a fusion technique that explicitly
deals with the properties of the resultant output image (to be generated through the
We have also discussed a method of band selection in Chap. 4 where a specific
subset of hyperspectral bands was selected from the input image based on the condi-
tional entropy measure. We have also observed that one can achieve almost a similar
fusion output by using a small fraction of hyperspectral data. If we choose the bands
properly, the resultant fused image obtained from this subset of the data provides a
very little degradation in the quality of the output when compared with the resultant
image obtained from the fusion of the entire dataset using the same fusion technique.
The remaining (discarded) bands, thus, contribute a very little amount of information
towards fusion as most of the independent information has already been captured.
In order to facilitate this band selection, the conditional entropy of a band of the
hyperspectral data given the set of already selected bands has been calculated.
The idea of the output-based fusion motivates us to develop an alternate band
selection method that selects a subset of bands depending upon whether the fused
image obtained by inclusion of a particular band to the subset will be significantly
different from the fused image obtained without inclusion of that band to the sub-
set. In this chapter, we shall again discuss the band selection method, but from an