Investigation of diversity and accuracy in ensemble of classifiers using Bayesian decision rules

Multiple Classifier System (MCS) has attracted increasing interest in the field of pattern recognition and machine learning where this technique has also been introduced in remote sensing. The importance of classifier diversity in MCS has been raised recently; nevertheless, only a few of the researches have been studied in land cover classification problem. In this paper, a SPOT IV satellite image covering the Hong Kong Island and Kowloon Peninsula with six land cover classes were classified with four base classifiers: Minimum Distance Classifier, Maximum Likelihood Classifier, Mahalanobis Classifier and K-Nearest Neighbor Classifier. Same training and testing data sets were applied throughout the experiments and five Bayesian decision rules, including product rule, sum rule, max rule, min rule, and median rule, were utilized to construct different ensemble of classifiers. Performance of MCS was measured using the overall accuracy and kappa statistics, and three statistical tests including McNemarpsilas test, Cochranpsilas Q test and F-test were introduced to examine the dependence of the classification results. The experimental comparison reveals that (i) increasing the number of base classifiers may not improve the overall accuracy in MCS, (ii) significant diversity in base classifiers cannot enhance the overall performance and vice versa. These findings are noted with the condition in using the same data set and the same training set.