Abstract
Proposed is a new type of discriminant function for classification of multispectral images. This uses both the mean and dispersion in each spectral channel obtained from a very small image area (subimage). The function is based on Neyman's ‘smooth’ test, which is a nonparametric test of a distribution model. Its performance is compared with those of one based on Kolmogorov Smirnov test and other discriminant functions by real multispectral image data in terms of classification accuracy and robustness for variation of training data. The comparison shows that
(1) this is quite efficient when variances of the data are largely different from one category to another, and that
(2) this is robust for variation of training data, but not very robust for a bias of mean caused by samples not representing the category.
