Abstract: The request for efficient compression of hyperspectral images becomes pressing. A cluster-based lossless compression algorithm for hyperspectral images is presented. Because the spectral correlation differs in different bands, spectral band grouping algorithm is introduced to divide hyperspectral images into groups according to the correlation between each adjacent bands. The important bands which contain large useful information can be determined by using the adaptive band selection algorithm, on which k-means clustering is carried out according to the spectral vectors. The current band is predicted by using several preceding bands. For each pixel which belongs to a certain cluster, some causal neighboring pixels which have been coded are trained to get the optimal predictive coefficients. The reference bands are compressed by JPEG-LS standard while the final predictive errors are coded by Golomb-Rice. Experimental results show that the proposed methods produce competitive results when compared with other state-of-the-art algorithms.