Hyperspectral subpixel unmixing via an integrative framework

ABSTRACT In hyperspectral applications, spectral unmixing (SU) is an important technology to obtain the endmembers and the fractional land covers. Spectral variability, outliers, and nonlinearity are three challenging issues, causing SU to extract endmembers and corresponding abundance maps inaccurately. However, in view of the complexity of the three issues, to tackle them at once is difficult and intractable. In this paper, all the aforementioned issues are advocated to process together by a powerful integrative SU framework, where a hyper-manifold learning approach via a sparsity-constrained dual is exploited. In the proposed integrative SU framework, heterogeneous and homogeneous information is explored by dividing the hyperspectral data (HD) into a series of sub-blocks, hyper-Laplacian-based graph is employed to address the nonlinearity, and spectral variability is controlled by a scaled matrix. Moreover, the interferences of the outliers are handled by the augmented correntropy-induced metric (ACIM), where the rare endmembers are separated from the abrupt anomalies via decomposing the HD sets in a low-rank structure and constraining the sparsity on the corresponding residual term. The experimental results on several popularly used real HD sets indicate the superior performances of the proposed approach.