Abstract: The traditional spectral unmixing algorithm considers that each pixel contains all the endmembers extracted from the image, which does not conform to the actual situation. In fact, most of the mixed pixels in the image are only mixed by a small number of endmembers. Because of the influences of endmember extraction precision and noise, if all endmembers are used in spectral unmixing, it will make the abundances of the endmembers which are not involved in the mixed pixel are not zero, the spectral unmixing results have large errors. Because most of the mixed pixels are located at the junction of different ground objects, this paper proposes a method to select the optimal endmember subset of mixed pixels by utilizing the spatial information of the image. Using a spatial structure element, this method starts to search pure pixel spectrum from the adjacent domains of the mixed pixels, then compares the searched pure pixel spectrum with the previously extracted image endmembers to determine the endmember subset of the mixed pixels. According to the variation of RMSE, the size of the structural element is gradually expanded, and the search scope is constantly adjusted until the optimal endmember set is obtained. The experimental results of the simulated data and the real data show that the proposed method has a relatively better spectral unmixing effect compared with the traditional spectral unmixing method using all endmembers.

Key words: hyperspectral image, mixed pixel, selective endmember