Double Weighted Sparse Nonnegative Tensor Factorization for Hyperspectral Unmixing

ABSTRACT

A variety of unmixing methods offered fruitful solutions for extracting endmembers and estimating abundances. Recently, a matrix-vector nonnegative tensor factorization (MV-NTF) unmixing method was proposed. Compared with nonnegative matrix factorization (NMF), NTF avoids the conversion of hyperspectral data from 3-D to 2-D, thereby preserving the intrinsic structure information. Nevertheless, MV-NTF ignores local spatial information owing to dealing with data as a whole. Thus, in this letter, to make the most of spatial information and abundance sparsity, a new double weighted sparse NTF (DWSNTF) unmixing method is proposed. Under the MV-NTF framework, a double weighted $L_1$ regularizer is firstly utilized to characterize more precise and sparse abundance maps. One weight acts on single pixel to promote the sparsity of solution, while the other weight exploits the local spatial information to conserve more details and prevent oversmoothness. In addition, all weights are stacked into a weight tensor to fit the higher-dimensional factorization and facilitate optimization. Experimental results on both synthetic and real data demonstrate the validity and superiority of our proposed method against the state-of-the-art methods.

Additional information

Funding

This work was funded in part by the National Natural Science Foundation of China under Grant [61871335] and [61801404] and in part by the Fundamental Research Funds for the Central Universities under Grant [2682020XG02] and Grant [2682020ZT35].