Endmember independence constrained hyperspectral unmixing via nonnegative tensor factorization

• Published in: Knowledge-based systems Vol. 216; p. 106657
• Main Authors: Wang, Jin-Ju, Wang, Ding-Cheng, Huang, Ting-Zhu, Huang, Jie, Zhao, Xi-Le, Deng, Liang-Jian
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 15.03.2021; Elsevier Science Ltd
• Subjects: Algorithms; Constraints; Correlation; Endmember independence constraint; Factorization; Hyperspectral imaging; Hyperspectral unmixing; Low-rankness; Mathematical analysis; Nonnegative tensor factorization (NTF); Spectra; Spectral signatures; Tensors; Hyperspectral unmixing; Low-rankness; Nonnegative tensor factorization (NTF); Endmember independence constraint
• ISSN: 0950-7051; 1872-7409
• DOI: 10.1016/j.knosys.2020.106657

Hyperspectral unmixing is an essential step for the application of hyperspectral images (HSIs), which estimates endmembers and their corresponding abundances. In recent decades, nonnegative matrix factorization (NMF) and nonnegative tensor factorization (NTF) have been widely exploited for hyperspectral unmixing. To improve the unmixing performance, various constraints have been applied in many NMF-based and NTF-based methods. Though many regularizations are used to describe abundances’ properties, less attention is paid to endmember signatures. Notice that, endmember information is important for obtaining accurate estimated endmembers from the highly correlated spectral signatures in HSIs. Thus, constraints on both endmembers and abundances are expected to make spectral signatures separated adequately. In this paper, we propose a new NTF-based model, termed as endmember independence constrained hyperspectral unmixing via NTF (EIC-NTF). It aims to mitigate the impact of high correlation among spectral signatures from endmembers and abundances. For endmember estimation, we introduce an endmember independence constraint to avoid obtaining similar endmembers estimations. For abundance estimation, we exploit the low-rankness in abundance maps to describe the spatial correlation of mixed pixels lying in homogeneous regions of HSIs. We solve the proposed model under the augmented multiplicative update framework. Experimental results on both synthetic and real hyperspectral data demonstrate that the proposed algorithm is effective for hyperspectral unmixing.