SVM-Based Unmixing-to-Classification Conversion for Hyperspectral Abundance Quantification

Need for a priori knowledge of the components comprising each pixel in a scene has set the endmember determination, rather than the endmember abundance quantification, as the primary focus of many unmixing approaches. In the absence of the information about the pure signatures present in an image scene, which is often the case, the mean spectra of the pixel vectors, directly extracted from the scene, are usually used as the pure signatures' spectra. This approach which is mathematically optimized for unmixing problems with a priori known information ignores some statistical properties of the extracted samples and leads to a suboptimal solution for real situations. This paper proposes a novel learning-based unmixing-to-classification conversion model to treat the abundance quantification task as a classification problem. Support vector machine, as an efficient classifier, is used to realize this model. It exploits the statistical nature (endmember spectral variability) of the extracted endmember representatives from the hyperspectral scene, rather than solving the problem according to the ideal model in which only the mean spectra of each training sample set is used. Several experiments are carried out on simulated and real hyperspectral images. The obtained results validate the high performance of the proposed technique in abundance quantification which is a key subpixel information detection capability.