Small class classification for hyperspectral remote sensing data

A simplified version of the maximum likelihood classification algorithm is presented in which the conventional assumption of class normality is modified by introducing a principal components transformation, allowing the correlation between the bands to be ignored and thus the transformed bands to be considered independently. This results in only the diagonal elements in the covariance matrix being taken into account, such that the class signature becomes the mean vector and variance vector. Therefore the number of training pixels per class is linked with the single band case, ie. It is reduced to below 100, from which training on small classes can benefit. The discriminant function then becomes the sum of the logarithmic discriminant values of each band. Data recorded by AVIRIS has been used to test the proposed method, showing the small class classification to be feasible with reduced classification time and high classification accuracy.