The classification and de-aliasing methods with respect to multi-spectra and hyper-spectra have been widely studied in recent years. And both K-mean clustering algorithm and spectral similarity algorithm are familiar classification methods. The present paper improved the K-mean clustering algorithm by using spectral similarity match algorithm to perform a new spectral classification algorithm. Two spectra with the farthest distance first were chosen as reference spectra. The Euclidean distance method or spectral angle cosine method then were used to classify data cube on the basis of the two reference spectra, and delete the spectra which belongs to the two reference spectra. The rest data cube was used to perform new classification according to a third spectrum, which is the farthest distance or the biggest angle one corresponding to the two reference spectra. Multi-spectral data cube was applied in the experimental test. The results of K-mean clustering classification by ENVI, compared with simulation results of the improved K-mean algorithm and the spectral angle cosine method, demonstrated that the latter two classify two air bubbles explicitly and effectively, and the improved K-mean algorithm classifies backgrounds better, especially the Euclidean distance method can classify the backgrounds integrally.