Abstract: Band selection from multispectral or hyperspectral image data is an effective method to remove redundancy among bands and thus reduce dimension. An efficient algorithm using divergence based class-within principal component analysis (PCA) and analysis of corresponding coefficients is proposed. At first, the covariance of each class is diagonalized through PCA transforms on class data respectively, and then the divergence only depends on the summation of individual feature separability of transformed bands. Secondly, after an analysis of corresponding PCA transform coefficients, the candidate bands, original bands essential to classification, are determined by majority vote. At last, the final band subset is obtained by analyzing the dependency and divergence of bands in every subset generated according to the correlations of original band in candidates. Compared with sequential forward selection, the proposed method reduces the computation complexity, and encouraging results have been shown by experiments with an Airborne Visible/InfraRed Imaging Spectrometer (AVIRIS) data set.