Ensemble classifier based training data refinement technique for classification of remotely sensed optical images

In this study, an ensemble classifier based method is investigated to detect and discard mislabelled training samples from training set. To decide, whether to discard a sample from the training set, majority voting is considered for an ensemble constructed using kNN, RBF neural network and SVM classifier which are diverse in their decision making. Further, this method is compared with conventional statistical anomaly detection based filter and multi-objective Genetic Algorithm (GA) based filters, in terms of its ability to handle the number of mislabeled samples present in the training set. The performance of all the filters is tested at different levels of mislabelling. At 95% confidence level, the ensemble classifier based filter is the best among all the implemented filters. It can detect and remove mislabeled training samples even if 40% training samples are mislabeled.