Noise removal in multichannel image data by a parametric maximum noise fraction estimator

Some approaches to noise removal in multispectral imagery are presented. The primary contribution of the present work is the establishment of several ways of estimating the noise covariance matrix from image data and a comparison of the noise separation performances. A case study with Landsat MSS data demonstrates that the principal components are not sorted correctly in terms of visual image quality, whereas the minimum/maximum autocorrelation factors and the maximum noise fractions (MAFs) are.

A case study with Landsat TM data shows an ordering which is consistent with the spatial wavelength in the components. The case studies indicate that a better noise separation is attained when using more complex noise models than the simple model implied by MAF analysis. (L.M.)