Interclass Separability Factorization for Hyperspectral Anti-Interference Unmixing

In the field of hyperspectral unmixing, to decompose mixed pixels into pure substances and their corresponding fractional abundances, the (hyper)graph constrained (D)(NMF) is typically applied under the guidance of the spectral information, as a result, the unmixing accuracy is highly dependent on the construction method of the (hyper)graph and how the spectral information is exploited. In addition, the (hyper)graph constrained (D)(NMF) models usually trapped in the problem of over-smoothness, that is, the pixels belonging to different substances and located in the edges are inevitably distorted. In this paper, a weighted NMF is constructed based on the spectral similarity between the bands, which is exploited from the physical characteristics of the hyperspectral image (HSI) and the corresponding interferences, respectively. Subsequently, a multi-higher-order graph is pursued and introduced to exploit the local smoothness of the abundances. This approach ensures that the edge information of each abundance map is expected to be preserved, and inter-class pixels can be successfully separated. Finally, an effective unmixing model called multi-graph regularized nonnegative matrix weighted factorization (MGNMWF) is constructed. The proposed model is experimented on both synthetic and real datasets. The experimental results show that the proposed model outperforms both shallow and deep unmixing models.