In order to solve the distribution and content of each substance in hyperspectral images, we introduce the hyperspectral classification into the endmember extraction to propose a new endmember extraction method. Firstly, the number of endmembers is determined by the virtual dimension. And the thought of hyperspectral unsupervised classification is performed by K-means clustering algorithm which classifies each pixel into classifications. Then the pixel with the largest spectral value is extracted from each kind of class. According to the theory of simplex, the pixels of hyperspectral image are used to form a simplex in the high-dimensional space, and the vertexes of the largest simplex are the endmembers which are extracted. The simulation and real data have shown that this method of endmember extraction has the advantages of high efficiency and accuracy compared with the traditional method.