A comparison of criteria for decision fusion and parameter estimation in statistical multisensor image classification

We study two related topics in decision fusion for multisensor image classification. The first topic is the use of a weighted logarithmic opinion pool compared to the statistical product combination rule. The performance is compared on three data sets. The second topic is related to different criteria for parameter estimation for a statistical fusion model. We propose an alternative criterion for estimation of the mean vector and the covariance matrix of a Gaussian model based on minimizing the number of misclassified training samples and compare the performance of this to the traditional maximum likelihood approach.