Adaptive tree-structured subspace classification of hyperspectral images

We present a scheme called “adaptive tree-structured subspace classification” (ATSC) for unsupervised classification of hyperspectral images. This novel approach is based on a self-organizing mechanism which results in adaptive subspace representation for each class. It integrates the process of feature extraction and classification under the framework of subspace pattern recognition. ATSC differs from normal subspace classification methods in that it can proceed without a predetermined number of class or trial-error iteration. This adaptivity is realized by introducing a tree structured dual subspace division algorithm in a competitive learning manner. The scheme is tested and verified on AVIRIS images