Because of their rich and complex texture, multi-band and high spectral resolution, hyperspectral images are widely used in military, marine, agriculture and other aspects; however, their huge data also brings many troubles. Currently, there’re almost three types of the hyper-spectral image compression methods: the method based on transform, prediction and vector quantization, among which, the method based on prediction has the advantages such as easy implement and higher compression ratio, then it’s widely used. In the present paper, features of hyperspectral images are analyzed first in detail, then inter-spectral prediction algorithm is improved, and a new scheme which consists of subspace partition and multi-inter-spectral prediction is proposed, and finally, the experiments show that the program is effective.