The technique of Multiple Classifier Systems (MCSs), which is a kind of decision-level information fusion,has fast become popular among researchers to fuse multiple classification outputs for better classification accuracy. In MCSs, there exist various kinds of uncertainties such as the ambiguity of the output of individual member classifier and the inconsistency among outputs of member classifiers. In this paper, we model the uncertainties inMCSs based on the theory of belief functions. The outputs of member classifiers are modeled using belief functions. A new measure of diversity in member classifiers is established using the distance of evidence, and the fusion rule adopted for MCSs is Demspter’s rule of combination. The construction of MCSs based on the proposed diversity measure is a dynamic procedure and can achieve better performance than using existing diversity measures. Experimental results and related analyses show that our proposed measure and approach are rational and effective.