Comparing the classification performances of supervised classifiers with balanced and imbalanced SAR data sets

Abstract

In this study, the classification accuracies of four different classification methods with two balanced and two imbalanced data sets for the classification of Sentinel-1B SAR (Synthetic Aperture Radar) data were comparatively evaluated and the impacts of training data sets into the accuracy were investigated. In some circumstances, it is possible to collect high number of ground truth samples for some classes however not possible for some other classes which are represented by less number of ground truth samples. In such cases, the imbalanced data set is the issue. Supervised classifiers, by its nature, employ many different input parameters in consideration of the decision surface separating the two classes. More than the classification model itself, purity, size and allocation of ground truth samples as well as the adaptation between the training data and adopted classifier are of key importance in accuracy of image classification. In our study, two parametric (Naive Bayes and Linear Discriminant Analysis) and two non-parametric (Support Vector Machines and Random Forests) supervised classification methods were implemented. Our experimental results demonstrated that there were not any significant change in classification accuracies of parametric classifiers and support vector machines however an increase in classification accuracy of random forest with imbalanced dataset. Furthermore, highest classification accuracy of this study (89.94%) was obtained by Support Vector Machines classification.