A Novel Endmember Bundle Extraction Framework for Capturing Endmember Variability by Dynamic Optimization

• Published in: IEEE transactions on geoscience and remote sensing Vol. 62; pp. 1 - 17
• Main Authors: Liu, Rong, Lei, Cong, Xie, Linfu, Qin, Xiaoqiong
• Format: Journal Article
• Language: English
• Published: New York The Institute of Electrical and Electronics Engineers, Inc. (IEEE) 2024
• Subjects: Algorithms; Optimization; Particle swarm optimization; Pixels; Root-mean-square errors; Variability
• ISSN: 0196-2892; 1558-0644
• DOI: 10.1109/TGRS.2024.3354046

The spectral variability problem is a big challenge in hyperspectral unmixing. Endmember bundles have been used to address the spectral variability problem by adopting a bundle of endmember spectra to represent one kind of endmember class. Existing endmember bundle extraction (EBE) algorithms mainly rely on the convex geometry assumption and integrate endmembers from image subsets as endmember bundles. On the one hand, they suffer from high risk of bad performance for real hyperspectral scene where the convex geometry assumption is not satisfied. On the other hand, endmember variabilities within image subsets are neglected, which may lose representative endmembers. In this article, we propose a novel EBE framework to capture endmember variability by introducing a dynamic optimization mechanism. Endmember bundles are obtained by dynamically minimizing the root-mean-square error (RMSE) between original pixels and reconstructed pixels through an iteration process, and a particle swarm optimization (PSO) method is introduced to find the optimal endmember combination in each iteration. The proposed EBE framework imposes no assumption on the hyperspectral data distribution and has great potential to be used in complex hyperspectral scenes. Experimental results on two real hyperspectral datasets demonstrate that the proposed algorithm is able to obtain endmember bundles that well express the spectral variability, and the performance of the proposed algorithm is competitive with the state-of-the-art algorithms.