Twin support vector machine-based hyperspectral unmixing and its uncertainty analysis

Abstract. In consideration of within-class endmember variability, it is realistic to use multiple endmembers to model a pure class. We propose an advanced multi-endmember unmixing algorithm based on twin support vector machines (UTSVM), which derives the abundances based on the distances from the mixed pixels to each classification hyperplane. Unmixing uncertainty, an issue often neglected in multi-endmember unmixing, is also analyzed quantitatively for UTSVM. Two types of unmixing uncertainty, abundance overlap (i.e., different mixed pixels have the same abundances) and model overlap (i.e., one mixed pixel may be unmixed into different abundances), are introduced. Abundance overlap angle and abundance variability scale (AVS) are defined as two uncertainty indexes to measure abundance overlap and model overlap, respectively. The relationship between within-class endmember variability and unmixing uncertainty is discussed. When the unmixing uncertainty is high, we propose to use the mean value of abundances within AVS as the estimation of abundance to obtain the best compromised results. Experimental results show the feasibility and effectiveness of our study.