Covariance matrix matching for multi-spectral image classification

This paper introduces two new statistics for use in classifying multispectral satellite images where classes are characterised by multispectral texture and are not easily separated by conventional algorithms. We provide motivation for their introduction and discuss some of their properties. Empirical results are presented to support the hypothesis that the covariance matrix of a pixel´s neighbourhood contains useful information when used for the classification of landcover in multispectral images. We speculate that these statistics could also provide an efficient means of matching multicoloured objects in computer vision problems, and in particular avoid some problems with lack of colour constancy