Efficient and accurate linear spectral unmixing

Björn Labitzke, Andreas Kolb

Abstract: The techniques of multi- and hyperspectral imaging have gained a growing attention in recent years. This is mostly due to their potential to provide rich information that can be used to improve material classification or product quality assessment. Linear spectral unmixing is a standard approach in hyperspectral data analysis. Based on the assumption that a spectral dataset can be expressed as a linear combination of constituent spectra, it is an important task to estimate the necessary coefficients. In the case that non-negativity of the coefficients is enforced, the calculation of the coefficients can be very time consuming. In this paper, we propose an GPU-based approach that efficiently and accurately computes the coefficients for linear spectral unmixing. Our approach is based on the orthogonal subspace projection technique and further can be combined with the image space reconstruction algorithm (ISRA) in order to improve the results in terms of accuracy and performance. We present detailed results of our proposed approach in comparison to ISRA for the domain of remote sensing.