An Improved MNF Transform Algorithm on Hyperspectral Images with Complex Mixing Ground Objects

Transforms on the spectral domain are often used to compress and extract information of sample features for the supervised classification on hyperspectral images. On condition that the ground samples are all in mass each, there are more obvious advantages on sample classification with obscure spectral features when using Maximum Noise Fraction (MNF) transform, or named two cascaded principal components analysis method, compared with Principle Components (PC) image transform. But through the experiments in this article below, it is proved that the classification accuracy will decrease obviously using MNF method when the samples of different ground object classes are complexly mixed. Then the article illustrates the rules and reasons for these influences. Aiming at this problem of the low classification accuracy when the ground samples are too scattered, an improve method of estimating image Noise Covariance Matrix (CM) of hyperspectral images is introduced to enhance the feature extraction effect of the MNF method. The followed experiment shows that the classification accuracy is increased effectively when using it compare with classic MNF method.