Unsupervised clustering and spectral unmixing for feature extraction prior to supervised classification of hyperspectral images

Classification and spectral unmixing are two very important tasks for hyperspectral data exploitation. Although many studies exist in both areas, the combined use of both approaches has not been widely explored in the literature. Since hyperspectral images are generally dominated by mixed pixels, spectral unmixing can particularly provide a useful source of information for classification purposes. In previous work, we have demonstrated that spectral unmixing can be used as an effective approach for feature extraction prior to supervised classification of hyperspectral data using support vector machines (SVMs). Unmixing-based features do not dramatically improve classification accuracies with regards to features provided by classic techniques such as the minimum noise fraction (MNF), but they can provide a better characterization of small classes. Also, these features are potentially easier to interpret due to their physical meaning (in spectral unmixing, the features represent the abundances of real materials present in the scene). In this paper, we develop a new strategy for feature extraction prior to supervised classification of hyperspectral images. The proposed method first performs unsupervised multidimensional clustering on the original hyperspectral image to implicitly include spatial information in the process. The cluster centres are then used as representative spectral signatures for a subsequent (partial) unmixing process, and the resulting features are used as inputs to a standard (supervised) classification process. The proposed strategy is compared to other classic and unmixing feature extraction methods presented in the literature. Our experiments, conducted with several reference hyperspectral images widely used for classification purposes, reveal the effectiveness of the proposed approach.