Expert-guided fuzzy/holistic hyperspectral image classification

A system is built whose purpose is to classify hyperspectral imagery. In this instance, such classification refers to the task of assigning to each pixel in an hyperspectral image a fuzzy truth value for membership in one of a number of preselected endmember classes. The classifier makes these assignments pixel by pixel, completely independently of the fuzzy truth assignments of neighboring pixels. The classifier works in two stages. First, an holistic classifier, utilizing a generalized spectral angle mapper, is run over the data to search for image pixels that look "something like" the target endmember pixels. Then, at each pixel so identified, spectral features are extracted from the hyperspectrum comprising that image pixel over some subrange of the data which is chosen by experience. Fuzzy truth values are assigned to each of these features, and a suite of fuzzy inference rules used to determine a composite fuzzy membership of each pixel in the class of endmembers. The final classification value is a composite of the truth values that result from the holistic classifier and the feature-based classifier. The system also puts out features selected by the user, so that the user can check to see if the pixels classified as a particular mineral in fact possess some relevant feature. In this system, the holistic classifier can be run generally, but the feature-based classifier must be driven by expert selection both of the features to use in the classification, and the relative weight each such feature should carry in the assignment of fuzzy truth. Each of these classifiers has complexity O(n), where n is the number of pixels in the image to be processed.