Spectral unmixing based on robust support vector machine

In traditional spectral unmixing method based on linear spectral mixing model (LSMM), unmixing accuracy is deteriorated both by the deficiency of the model and the participation of unrelated classes. Therefore, to solve this problem, a method with spatial information considered was proposed based on linear least square support vector machines (LLSSVM). A robust LLSSVM was constructed and then applied to spectral unmixing. Then pure pixels were determined and unmixing results were corrected with related classes selection for each mixing pixel. Ultimately LLSSVM was used again for spectral unmixing with only related classes considered. Simulation results indicate that the unmixing accuracy of the proposed method is improved by 10% higher than that of the traditional methods and is also high in efficiency.