Hyperspectral images are extremely susceptible to noise interferences which hinder mineral identification and quantification. Therefore, it is fundamental the use of techniques to minimize them. An efficient mathematical noise reducing procedure employs the Minimum Noise Fraction (MNF). The MNF procedure is derived from and similar to the Principal Component Analysis (APC). The main difference between them is that the MNF considers the noise while the PCA considers the data variation. It results in an ordering that reflects the image quality. The MNF transformation can be subdivided in four stages: a) obtaining a noise sample; b) formulation of a noise fraction index; c) implementation of a linear transformation function as PCA; and d) inversion of MNF considering only signs information. In the present work a methodology of MNF sequential steps is proposed using internal and external references of the image. As external reference it was used the dark reference collected during a flight simultaneously to the image acquisition. The internal reference was obtained by statistical techniques for the segmentation of the noise fraction in relation to the sign. Three types of noises can be defined by variance and correlation: a) uncorrelated noises with equal variance in all bands, b) highly correlated noises, and c) Noises with Unknown Covariance. Matrix. For the AVIRIS images, a sequential procedure of MNF can be accomplished with two stages: the first one eliminates the noises of intern reference and the second the relative external reference to the instrumental noise. That procedure allows to distinguish different types of noises and to eliminate them sequentially and efficiently.
Remote sensing; image processing; hyperspectral; noise
