Unmixing method for hyperspectral data based on subspace method with learning process

Abstract

An unmixing method for hyperspectral data based on subspace method with learning process is proposed. Unmixing methods can be divided into three categories, inversion method, SVD (Singular Value Decomposition) based method, and subspace method. Although these methods works well if the distribution of the hyperspectral data in feature space can be represented as somewhat convex function, it is now allways true. Unmixing method proposed here does work even if the distribution is expressed with concave function because the method adjust the axis of subspace through a learning process. Through experimental studies with AVIRIS (Airborne Visible InfraRed Imaging Spectrometer) data, it is found that the proposed method achieves 16.3% of improvement of the unmixing accuracy in terms of root mean square error in comparison to the well known least square unmixing method. Also it is found that the proposed method shows 15.0% better accuracy in comparison to the subspace based unmixing method. The reasons for the improvement are clarified in a comprehensive manner with the simple example of the 3D feature space together with two categories.