EVALUATION OF MULTI CLASSIFIER

Abstract

Digital image classification techniques are applied to generate thematic maps from remote sensing data. For these classifications, the conventional parametric statistical classifiers, which have been applied in remote sensing for last two decades, are not appropriate, since they are limited to some extent where they cannot answer multivariate problems in data. Multiple classifiers, also known as classifier ensemble, have shown their ability to fuse together multiple classification outputs for better accuracy of classified image and hence improvement over conventional parametric statistical classification methods. Bagging and Boosting has been used in this study for implementation of multiple classifiers, among available multi classifiers. The factors that are used for creating a perfect multi classifier, is studied in this work theoretically. Neural network has been used as a base classifier in Bagging and Boosting algorithm as well as a single network that is used for comparisons, because neural network has been extensively applied in different domains and Bagging and Boosting would be very effective when they are applied on the classifiers which are unstable, that is which can change there result unevenly even for a slight change in input data. Here the performance of these Bagging and Boosting algorithms in terms of accuracies and sensitivity towards network parameters have been compared with a single network that is neural network by using a remote sensing dataset i.e. LISS-III image. Experimentally it is found that, the multiple classifiers outperform the single classifier in terms of overall accuracies and sensitivity to changes in network parameters. Overall accuracies of single network classifier i.e. Neural Network, Bagging and Boosting are 82.66, 86.66 and 86 respectively. It is also observed that sensitivity toward changes in network parameters also results in same and bagging and boosting has outperformed the single classifier. Bagging and Boosting algorithms has been implemented using neural networks as the classifier system, and to compare the results from both Bagging and Boosting with that of a single classifier. The work carried out here suggest that Bagging and Boosting ensembles have a higher accuracy rate than a single network, and are less sensitive to the power of the network used and to the amount of training that the networks have.