Target representation in hyperspectral images based on tensor block term decomposition

Combining imaging technology and spectroscopy in a unified system, hyperspectral data sets provide a powerful sense to discriminate the targets in a view. However, most of the traditional methods for HSIs concentrate on pixel/subpixel level targets and they only rely on spectral characteristic. On the one hand, the target composed of different materials may not be well-represented by only one kind of spectrum. Consequently, target recognition probability would be reduced or even failure. On the other hand, different targets might be composed of the same or similar materials, thus creating a source of false alarms. In this paper, a spectral-spatial representation for the target in hyperpsectral images (HSIs) is proposed, under the theory of tensor block term decomposition (BTD). As a consequence, the target is modeled by a set of spectrum-image terms. In each term, the spectrum indicates one kind of material of the target and the counter image corresponds to the spatial distribution of such spectrum. Both the spectral and spatial characteristics of the target are described for improving the effectiveness of hyperspectral target recognition technology. Experiments with both simulated and real HSI data sets reveal that the proposed method outperforms those spectral-based methods with better target discrimination.