Experiments in multimodality image classification and data fusion

Abstract

In this paper, we report the results of some experiments on image classification and data fusion of remote sensing images, as part of ongoing efforts at the CVIP to develop a general strategy for the analysis of multimodality imaging. Statistical and Fuzzy logic approaches have been employed in these experiments. In all, six different algorithms for image classification, and an image fusion algorithm have been implemented and evaluated on common data sets. These algorithms are: (1) Supervised Parametric Bayes Classifier; (2) Non-parametric Bayesian Classifier using the Parzen density estimate; (3) Maximum a posteriori classification using the k-nearest neighbors (k-NN) approach; (4) MAP Estimation using Markov random field modeling; (5) a Fuzzy logic approach; and (6) a novel discriminate function classifier. The AMP segmentation of the regions in the image has been implemented using the Iterated Conditional Mode (ICM) optimization method. This approach provided the best results, in terms of the minimum probability of error and best reliability. A novel decision fusion algorithm, based on the a priori class conditional probability, has been applied to the classifiers' output.