Space object material identification of hyperspectral data using nonnegative tensor factorization

Abstract

Among kinds of ways to improve the early-warning of a country, identifying the space object material in a better and faster way is an important and effective method. The hyperspectral image, which is a 3-D data cube and contains the spatial and spectral information of the interest objects, will play a more important role in identifying the space object material. However, the low spatial resolution of the hyperspectral remote sensing instrument makes the single pixel spectrum often mixed up several different materials' spectra, which is called mixed pixel. So it is a considerable question to decompose the mixed pixels into spectra of pure materials (called endmembers) and get their corresponding fractions (called abundances). Since a hyperspectral image can be seen as a 3-D tensor, nonnegative tensor factorization (NTF) algorithm based on tensor analysis can be introduced into the field of hyperspectral unmixing. However, random initialization, a classical way to initialize the NTF algorithm, causes a slow rate of convergence, which can be improved through other methods to initialize this algorithm. This paper selects the vertex component analysis (VCA) algorithm to initialize the NTF algorithm. In this way, a faster and better result is obtained, and furthermore, four simulated hyperspectral images dataset of 3-D model of Hubble Space Telescope with different spatial resolutions are processed by the improved algorithm in this paper, and good results are obtained.