An unsupervised hyperspectral image classification method simultaneously realizing the mixture of probabilistic PCA and clustering under the frame of EM algorithm is proposed. It is based on the fact that different class should have its own representative feature set, and it realizes feature extraction and classification in one step while preserving as much separability. It also possesses the advantages of PPCA model, which is more effective to high dimensional data processing. Applying the method to simulated data and real data shows that it can achieve better results compared with the method that applies PCA to all data without differentiation among classes.