A greedy algorithm for sparse unmixing

Abstract

Hyperspectral imaging sensors provide image data containing both spatial and detailed spectral information. However, due to low spatial resolution, the pixels in hyperspectral images are actually mixtures of the spectral signatures of the materials. Sparse unmixing assumes that these mixed pixels are sparse linear combinations of different material spectra which are in a spectral library. However, spectral libraries contain materials that have similar spectral characteristics. That's why greedy algorithms may choose incorrect material from the spectral library that is similar to the material in the mixed pixel in the first iteration. This leads to incorrect material selections in the subsequent iterations. In this study, an orthogonal matching pursuit (OMP) variant method is proposed to deal with this issue. Our proposed method is compared to OMP-Star and SunGP methods. Experiments on simulated and real hyperspectral data show that satisfactory results have been achieved by using the proposed method.