Every pixel in the super space is required by K-means algorithm to calculate Euclidean distance for clustering.When there are many class centers

this is a rather time consuming work.In this paper

an improved K-means clustering algorithm is presented to save initial clustering time by making initial division based on previous clustering results

to remain the stable relationship between classes

and to accelerate clustering process with more and more classes becoming stable by judging the centers nearest to the pixel.A new clustering lossless compression algorithm designed here can determine the best class number and the highest compression ratio by fully utilizing previous clustering results and converging quickly eliminating the inter-spectral redundancy and intra-intra-spectral redundancy through enhancing the intra-class pixel redundancy.The convergence of this algorithm and existence of the best parameters are also inferred by making a deep analysis of the probability distribution model of the residue data.Furthermore

the comparison with DPCM lossless compression algorithm in the entropy value of the probability distribution model and the experimental results show that this clustering algorithm is better than non-clustering compression algorithm.Several times clustering approach can forecast the best class number with the least entropy lossless compression.