Technical note Non-parametric test of overlap in multispectral classification

Abstract

Many strategies have been developed for the allocation of image pixels, or picture elements, to classes in order lo obtain segmentation of images. Difficulties are encountered when a pixel may be assigned to more than one class according to the information gleaned from training sets. The following question arises: is there real overlap or are the training sets spectrally discrete? A general algorithm is presented which quantifies the degree of overlap of classes defined by training sets drawn from an image. Test results show that age classes within a forest plantation are spectrally discrete, even though poor classification accuracies were obtained using conventional classifiers.