An adversarial learning mechanism for dealing with the class-imbalance problem in land-cover classification

Abstract

In this manuscript, a deep learning based approach has been investigated for land-cover classification capable of handling situations where there is an imbalance in the class-wise samples in the training set under a novel semi-supervised learning framework. This problem is persistent when the presence of some of the major land-cover classes affects the other recessive classes in a particular region and due to this the generated training set contains a fewer number of samples from the latter group of classes. Here, an adversarial auto-encoder has been used to generate synthetic samples from each of the minority classes so that each of the classes is well represented in the training set. A new training set has then been designed by including the ‘most confident’ artificial samples from the minority classes. This newly formed balanced training set is then shown to be more effective to classify the test samples as compared to the initially available imbalanced training set. The validation of the proposed approach has been performed using patterns collected from two multi-spectral satellite images captured by Ikonos-2 and landsat-8 satellites over different regions of India. The results show significant improvement in test class prediction when compared to that of other state-of-the-art imbalance handling schemes in land-cover classification.