Segmentation of hyperspectral images using local covariance matrices in eigenspace

In this work, segmentation of hyperspectral images by local covariance matrices in eigenspace has been proposed for getting high accuracy rates using unsupervised methods. Combination of both spectral and spatial features can increase the segmentation accuracy for hyperspectral images without groundtruth. Furthermore, changing from original data space to eigenspace via principal component analysis and its kernelized version and the calculation of covariance matrices in this new space can produce better results for different clustering methods. In the simulations, effects of local neighbors in the computation of covariance matrices in eigenspace were represented using four different clustering algorithms comparatively.