Multiple classifier systems for supervised remote sensing image classification based on dynamic classifier selection

In the literature, multiple classifier systems (MCSs) have proved to be a valuable approach to combining classifiers, and under some conditions MCSs are able to mimic ideal Bayesian labeling. This paper focuses on the family of MCSs based on dynamic classifier selection (DCS) and proposes a modification to dynamic classifier selection by local accuracy (DCS-LA). Experiments show that the proposed method outperform MCS strategies based on belief functions and the DCS-LA in terms of minimum and maximum class accuracies and kappa coefficient of agreement and is a valid alternative to majority voting. Moreover, the experiments show that MCSs based on the classification results of classifiers characterized by a low design complexity like maximum likelihood and nearest mean classifiers can yield accuracies that are quite comparable to those of highly optimized classifiers.