Learning A Common Subspace from Hyperspectral-Multispectral Correspondences

With a large amount of multispectral imagery available (e.g. Sentinel-2, Landsat-8), considerable attention has been paid to global multispectral landcover classification. There is, however, a typical bottleneck for further improving the performance of classification in the poor spectral information of multispectral data. On the contrary, hyperspectral data fails to be largely collected but is characterized by rich spectral information. To this end, we aim to learn a common subspace from hyperspectral-multispectral correspondences by simultaneously considering subspace learning and classification. Local manifold structure jointly constructed from different modalities is further embedded into the proposed framework. With the learned projections, the multispectral out-of-samples can be smoothly projected into the common subspace, which are expected to be better clarified. Extensive experiments on two HS-MS datasets where MS data sets are theoretically generated by their corresponding HS data, are performed to demonstrate the superiority and effectiveness of the proposed method in comparison with several state-of-the-art methods