Morphological scale-space for hyperspectral images and dimensionality exploration using tensor modeling

This paper proposes a framework to integrate spatial information into unsupervised feature extraction for hyperspectral images. In this approach a nonlinear scale-space representation using morphological levelings is formulated. In order to apply feature extraction, Tensor Principal Components are computed involving spatial and spectral information. The proposed method has shown significant gain over the conventional schemes used with real hyperspectral images. In addition, the proposed framework opens a wide field for future developments in which spatial information can be easily integrated into the feature extraction stage. Examples using real hyperspectral images with high spatial resolution showed excellent performance even with a low number of training samples.