A Novel Nonnegative Matrix Factorization Method for Hyperspectral Unmixing

摘要

In this paper, we propose a new algorithm integrating pure pixel identification into nonnegative matrix factorization(NMF) model to decompose the mixed pixels existing in hyperspectral imagery. The proposed algorithm employstraditional endmember identification algorithm to search for the pure pixel candidates, and then the principal componentanalysis is performed on the homogenous pixels which consist of the pure pixel candidates and its neighborhoods toidentify the endmembers existing in the real scene. Finally, the known-endmember-based NMF unmixing algorithm isused to generate the other unknown endmembers. The proposed algorithm retains the advantages of both pure pixelidentification method and NMF. Experimental results based on simulated and real data sets demonstrate the superiorityof the proposed algorithm with respect to other state-of-the-art approaches.