The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLII-3
https://doi.org/10.5194/isprs-archives-XLII-3-329-2018
https://doi.org/10.5194/isprs-archives-XLII-3-329-2018
30 Apr 2018
 | 30 Apr 2018

ENDMEMBER EXTRACTION OF HIGHLY MIXED DATA USING L1 SPARSITY-CONSTRAINED MULTILAYER NONNEGATIVE MATRIX FACTORIZATION

H. Fang, A. H. Li, H. X. Xu, T. Wang, K. Jiang, Y. Z. Su, and H. W. Chang

Keywords: Hyperspectral Imagery, Nonnegative Matrix Factorization, Multilayer Nonnegative Matrix Factorization, Sparsity Constraint, Endmember Extraction

Abstract. Due to the limited spatial resolution of remote hyperspectral sensors, pixels are usually highly mixed in the hyperspectral images. Endmember extraction refers to the process identifying the pure endmember signatures from the mixture, which is an important step towards the utilization of hyperspectral data. Nonnegative matrix factorization (NMF) is a widely used method of endmember extraction due to its effectiveness and convenience. While most NMF-based methods have single-layer structures, which may have difficulties in effectively learning the structures of highly mixed and complex data. On the other hand, multilayer algorithms have shown great advantages in learning data features and been widely studied in many fields. In this paper, we presented a L1 sparsityconstrained multilayer NMF method for endmember extraction of highly mixed data. Firstly, the multilayer NMF structure was obtained by unfolding NMF into a certain number of layers. In each layer, the abundance matrix was decomposed into the endmember matrix and abundance matrix of the next layer. Besides, to improve the performance of NMF, we incorporated sparsity constraints to the multilayer NMF model by adding a L1 regularizer of the abundance matrix to each layer. At last, a layer-wise optimization method based on NeNMF was proposed to train the multilayer NMF structure. Experiments were conducted on both synthetic data and real data. The results demonstrate that our proposed algorithm can achieve better results than several state-of-art approaches.