Abstract: In order to perform hyperspectral image clustering without the guidance of prior training samples’ label,a novel unsupervised extreme learning machine（ELM） block clustering algorithm for hyperspectral image was proposed.Firstly,the image was pre-clustered and the training samples were selected by using block strategy.Secondly,based on traditional spectral clustering algorithm,the ELM prediction mechanism was introduced.Then,the optimal output matrix of the ELM was calculated with the training samples.Finally,the feature map of the whole image was acquired by the optimized ELM,and clustering was achieved in the embedding space.Compared to traditional algorithm,six comparative experiments indicate that the proposed method can overcome the bottleneck of computational memory problem and achieve the higher precision clustering of large size image.