Unsupervised Hyperspectral Unmixing via Kernelized Correlations

In this work we derive a novel unsupervised scheme for unmixing hyperspectral pixels. Building on a novel combination of canonical correlations with unsupervised kernel learning, a novel unmixing technique is formulated that can perform even when data are nonlinearly mixed. Principal component analysis is utilized to set proper kernel parameters, while projected gradient descent techniques are used to obtain an estimate of the abundances, and corresponding endmembers. Numerical tests on real hyperspectral data advocate that the proposed method outperforms existing techniques, even in the presence of missing pixel entries.