In the past few years

there has been a growing interest in the development of automatic change detection techniques for the analysis of multi-temporal remote sensing images. This paper introduces a method for multivariate change detection

which is based on the canonical correlation analysis and the orthogonal transformation. Differing from traditional multivariate change detection schemes such as the principal component analysis （PCA）

this method takes two co-registered multivariate or multi-spectral satellite images covering the same geographic area typically acquired at different times as a whole random sample

and transforms two sets of random variables into a new set of random variates by using canonical transformation. To overcome the problem of lacking automatic techniques for discriminating the changed and unchanged pixels in the difference image

we propose an automatic technique based on the Bayes theory for the analysis of difference image. It assumes that the difference magnitudes comply with normal distributions. An automatic method for selection of the decision threshold that minimizes the overall change detection error probability is investigated. To perform an unsupervised estimation of the statistical terms that characterize these distributions

an iterative method based on the Expectation-Maximization （EM） algorithm is also presented. The experimental results show the fact that the presented method is exactly creditable and effective in multivariate change detection of remote sensing satellite data.