RANDOM-PROJECTION-BASED NONNEGATIVE LEAST SQUARES FOR HYPERSPECTRAL IMAGE UNMIXING

Nonnegative least squares, a state-of-the-art approach to end-member abundance estimation in the hyperspectral-unmixing problem, is coupled with random projection employed for dimensionality reduction. Both Hadamard- and Gaussian-based projections are considered. Experimental results reveal that random projections can significantly reduce data volume without detrimentally affecting the accuracy of the abundance estimation, with the Hadamard-based approach slightly outperforming its Gaussian counterpart.