Supervised Hyperspectral Image Classification Based on Spectral Unmixing and Geometrical Features

The spectral features of hyperspectral images, such as the spectrum at each pixel or the abundance maps of the endmembers, describe the material attributes of the structures. However, the spectrum on each pixel, which usually has hundreds of spectral bands, is redundant for classification task. In this paper, we firstly use spectral unmixing to reduce the dimensionality of the hyperspectal data in order to compute the abundance maps of the endmembers, since the number of endmembers in an image is much less than the number of the spectral bands. In addition, using only the spectral information, it is difficult to distinguish some classes. Moreover, it is impossible to separate objects made by the same material but with different semantic meanings. Some geometrical features are needed to separate such spectrally similar classes. In this paper, we introduce a new geometrical feature—the characteristic scales of structures—for the classification of hyperspectral images. With the help of the abundance maps obtained by spectral unmixing, we propose a method based on topographic map of images to estimate local scales of structures in hyperspectral images. The experiments show that using geometrical features actually improves the classification results, especially for the classes made by the same material but with different semantic meanings. When compared to the traditional contextual features (such as morphologB. Luo (B) · J. Chanussot Department of Image and Signal, GIPSA-Lab, Grenoble, France e-mail: [email protected] J. Chanussot e-mail: [email protected] ical profiles), the local scale provides satisfactory results without significantly increasing the feature dimension.