A robust subspace method for semiblind dictionary-aided hyperspectral unmixing

摘要

Recent development in semiblind dictionary-aided hyperspectral unmixing (HU) shows that a classical method in sensor array processing, namely, multiple signal classification (MUSIC), provides an effective way for endmember identification. However, MUSIC (and in fact, other dictionary-based sparse regression algorithms) assumes that there are no mismatches between the true endmember signatures and the dictionary spectral signatures, which may be violated in practice owing to reasons such as endmember variability and calibration errors. This paper presents a robust MUSIC method, wherein spectral signature mismatches are incorporated in the original MUSIC formulation to make the resulting algorithm robust. A computationally simple method is derived for the implementation of robust MUSIC. Simulation results show that robust MUSIC provides improved robustness against spectral signature mismatches than the original MUSIC.