Endmember Extraction Using Minimum Volume and Information Constraint Nonnegative Matrix Factorization

Abstract

Simplex volume is the most commonly used parameter for nonnegative matrix factorization (NMF)-based endmember estimation methods, and one of the most popular methods is the NMF method with minimum volume constraint (MVC-NMF). However, when outliers exist in the image, MVC-NMF tends to extract them as endmembers. In most cases, those outlier endmembers could be either physically meaningless or not representative enough for prevalent land covers. So how to extract prevalent land covers instead of outliers as endmembers is a very challenging question. In this letter, we propose a new NMF method with the dual constraints of simplex volume and information content, named the “minimum volume and information constraint NMF” (MIVC-NMF). The method is based on the following facts: when a real endmember is replaced by an outlier, it will cause some pixels containing the replaced endmember not to locate within the endmember hyperplane, and the overall information content contained in the endmember hyperplane will be reduced. The experimental results based on the simulated and real data show that the proposed method outperforms several other commonly used endmember extraction approaches.