Short introduction to the principal components transformation and the canonical discriminant analysis of multispectral data

This report reviews shortly two different algorithms for multispectral data analysis: the principal components transform (PC transform) and the kanonical discriminant analysis (KD analysis). These algorithms are widely used in multispectral and hyperspectral remote sensing for classification and identification of different objects, e.g. different types of crops, vegetation and minerals. Both the algorithms find a set of eigenvectors or filter functions which include the spectral features which are necessary for optimal classification but exclude redundant spectral data. While the PC transform treats the different spectral classes globally without concern to class structure in the data the KD analysis will maximize the ratio between the among-class variance and the within-class variance, by which class separation is optimized.