Virtual dimensionality estimation by double subspace projection for hyperspectral images

Virtual Dimensionality (VD) estimation is a key problem in feature/band selection and spectral mixture analysis of hyperspectral images. In this paper, a Double Subspace Projection (DSubP) based VD estimation algorithm is proposed. The pixel representation and image representation of a hyperspectral image are utilized to generate two subspaces according to the principal component analysis (PCA), respectively. When the dimensionality of these two subspaces exceeds VD of the hyperspectral image, both subspace projections show the same reconstruction performance. Therefore, VD can be estimated by judging the difference of reconstruction performance between DSubP. Both synthetic and real hyperspectral experiments demonstrate that the performance of the proposed DSubP based VD estimation algorithm outperforms that of the HFC and NWHFC based VD estimation algorithms.