Assessment of a Modified Version of the EM Algorithm for Remote Sensing Data Classification

This work aims to present an assessment of a modified version of the standard EM clustering algorithm for remote sensing data classification. As observing clusters with very similar mean vectors but differing only on the covariance structure is not natural for remote sensing objects, a modification was proposed to avoid keeping clusters whose centres are too close. Another modification were also proposed to improve the EM initialization by providing results of the well known K-means algorithm as seed points and to provide rules for decreasing the number of modes once a certain a priori cluster probability is very low. Experiments for classifying Quickbird high resolution images of an urban region were accomplished. It was observed that this modified EM algorithm presented the best agreement with a reference map ploted on the scene when compared with standard K-means and SOM results.