Unsupervised hyperspectral image classification using blind source separation

This paper presents an unsupervised classification algorithm for hyperspectral remotely sensed imagery based on blind source separation. Since the area covered by a single pixel in such an image is very large, the reflectance of a pixel is the mixture from all the materials resident in this area. A contrast function consisting of the mutual information minimization and orthogonality among the outputs, is defined to separate the assumed linear mixture so as to achieve soft classification. In order to reduce the computational complexity, a Neyman-Pearson detection theory based eigen-thresholding method is used to estimate the number of classes, followed by a band selection technique to select a smaller number of bands used in the learning algorithm. The preliminary result using an AVIRIS experiment demonstrates the feasibility of the proposed algorithm.