Evaluation of diversity measures for multiple classifier fusion by majority voting

Recently, decision fusion has shown great potential to increase classification accuracy beyond the level reached by individual classifiers. However, dependencies among classfiers' outputs strongly influence performance in a multiple classifier system (MCS) and thus have to be taken into account. In this paper, diversity between different classifiers used in remote sensing is assessed statistically. Several such measures are surveyed and evaluated on real benchmark remote-sensing datasets and using simulations.. The quality of a measure is assessed by the improvement in the accuracy of the multiple classifiers combined by the simple, yet efficient majority voting rule. Our experiments show that some diversity measures can indeed predict the performance of the fused classifiers, and thus should be considered on designing a MCS.