Impact of initialization on nonnegative matrix fraction for endmember extraction for hyperspectral imagery

Abstract

© 2016 IEEE. Nonnegative matrix factorization (NMF) has been widely applied to hyperspectral unmixing in these years. The solution of NMF highly depends on the initialization of endmembers and abundance fractions. This paper mainly addresses one issue: how to set appropriate initial endmembers and abundance for NMF. It is found that when the initial endmembers are close to real endmembers and the initial abundance is calculated by fully constrained least squares algorithm, NMF converges very slowly. However, the converging rate will significantly increase if a disturbance is added to the initial abundance. Experiments with both simulated hyperspectral data show the effectiveness of the proposed method.