Automatic identification and extraction of endmember from hyperspectral imagery by iterative unmixing

Current algorithms of endmember extraction basically need manually determining the number of endmembers, which is not conducive to automatically process. The paper puts forward iterative algorithm for automatic identification and extraction of endmember. First, we obtain the similarity threshold among pixels by statistical analysis, and determine the criterion of candidate endmembers. Then, the internal and external correlation judgments of candidate endmembers are done, and ill-conditioned matrix to circumvent judgment on endmember spectral set is conducted. Finally, the criterion of candidate endmembers is the end of the iterative conditions. When the hyperspectral image contains no candidate endmembers, the endmember spectral set is got and the numbers of endmembers are determined. Experiments show the effectiveness of this method, by which the error risk of sequential endmember extraction algorithm can be avoided, and the degree of automation is improved.