Abstract: Two nuisances will be encountered in classifying the multi-spectral data set using the EM algorithm. The first is the singularity of the variance–covariance matrix,which will lead to the failure of the computation; the second is the sensitivity to the initial values selected at random ,which will result in different final classification results. This paper suggests that the difference of classes is emphasized by logarithmizing the original data, and then apply the principal component transformation to the logarithmized data. The number of principal components for the EM algorithm is determined according to the contribution rate of more than 85% of all principal components. This not only removes the singularity but weakens the noise; The histogram of the first principal component of the logarithmized data reflects sufficiently the information of the class difference, from which the initial label for multi-dimensional EM algorithm can be efficiently determined, and the sensitivity of the initial value selected at random can be avoided. The experiment shows that the proposed strategy is better than the general EM algorithm and K-means.