An improved EM algorithm for remote sensing classification

The use of a general EM (expectation-maximization) algorithm in multi-spectral image classification is known to cause two problems: singularity of the variance-covariance matrix and sensitivity of randomly selected initial values. The former causes computation failure; the latter produces unstable classification results. This paper proposes a modified approach to resolve these defects. First, a modification is proposed to determine reliable parameters for the EM algorithm based on a k-means algorithm with initial centers obtained from the density function of the first principal component, which avoids the selection of initial centers at random. A second modification uses the principal component transformation of the image to obtain a set of uncorrelated data. The number of principal components as the input of the EM algorithm is determined by the principal contribution rate. In this way, the modification can not only remove singularity but also weaken noise. Experimental results obtained from two sets of remote sensing images acquired by two different sensors confirm the validity of the proposed approach.