A growing classifier applied to partially labeled Landsat images

A method to automatically generate a Gaussian mixture classifier is presented. The growing process consist of iterative addition of a new Gaussian mixture. Every iteration is divided into two sequential phases: first, the likelihood of the data under the current configuration is maximized by means of the EM algorithm and then a new Gaussian mixture is added in the class that need it most in terms of a discriminative rule. Growth control is imposed by a complexity penalizing term and by a discriminative condition. After the growing process is finished a combined re-estimation using labeled and unlabeled data is performed. We report the results on some artificially generated examples and on terrain classification over a Landsat-TM image using different restrictions for the covariance matrix of the mixtures