Hyperspectral unmixing using a novel conversion model

In absence of prior knowledge of the pure signatures (endmembers) existing in a remotely sensed image which is often the case, the mean spectra of the pixel vectors directly extracted from the image scene are usually used in unmixing problems. This approach ignores some important statistical properties of the extracted samples, thus, leads to suboptimal solutions. This paper proposes a novel method for unmixing of hyperspectral imagery through a classification model. It first transforms the unmixing problem into a classification task where the abundances of endmembers can be estimated with the help of a set of artificially made classes using known endmember compositions. Support vector machine, as an efficient classifier, is used to realize this model. The proposed method exploits the statistical nature of the extracted endmember representatives. Experiments on hyperspectral images validate the high performance of the proposed method in unmixing which is a key subpixel information detection technique.