Assessing reliability of classification in the most informative spectral regions of hyperspectral images

Reliability analysis is usually applied to evaluate classification procedures with different classes. In this research, we have applied the analysis to two different band sets to find out which one is more reliable. These band sets provide the most informative spectral regions covered by hyperspectral images. The informative regions are identified by minimizing two dependency measures between bands: correlation coefficient and normalized mutual information. The implementations are done by a newly developed top-down method named Spectral Region Splitting (SRS) resulting in two sets of bands which are almost identical at critical spectral regions. A reliability analysis based on the thresholding technique of the two sets of bands was performed. A technique was applied to discard those pixels that are not correctly classified at the given confidence level. The results show that the informative spectral regions selected by normalized mutual information was more reliable.