Superpixel-guided multifeature tensor for hyperspectral image classification with limited training samples

Supervised hyperspectral image (HSI) classification is challenged by the deficiency of labeled samples. The spatial correlation and multifeature have been proved to be very helpful for HSI classification. Thanks to the multiway structure, the tensor can express a sample by its spatial correlation and multifeature. However, integrating heterogeneous features and spatial correlation into a tensor leads to very high data dimensionality, which is fatal for limited training samples case. In addition, most multifeature methods devote to maximizing the agreements on heterogeneous features, while the inherent structures of each specific feature are not noticed. To address these problems, in this paper, we propose a superpixel-guided multifeature tensor (SPGMF) method for HSI classification which associates superpixel (SP) with multifeature through tensor, hence, solving the problem of limited training samples. Specifically, SPs guide to expanding training set as well as capturing local similarity. Subsequently, multifeature pixels from a SP are transformed into a latent space and stacked into a tensor, as a result, SPGMF not only captures the local similarity of HSI but also controls the dimensionality increment. Furthermore, a low-rank and sparse tensor decomposition regularized by multigraph is proposed, so that the consistency of multifeature is maximized and the local structure of a specific feature is preserved. Extensive experiments on three benchmark HSIs demonstrate the effectiveness and superiority of the proposed SPGMF, particularly with very limited training samples.