Selective Data Augmentation Approach for Remote Sensing Scene Classification

Abstract

Scene classification in remote sensing (RS) images is an issue that attracted a lot of researchers' attention recently. Using CNN for scene classification has been investigated in depth. One difficulty in using CNN models in remote sensing is the limited amount of data. Data augmentation techniques have been shown to provide one solution to this problem, yet, few works have investigated these techniques in their methods. In this work, our main contribution is presenting a novel method for selective data augmentation in remote sensing. The proposed selective augmentation method tries to optimize the way we augment the training set. We do that by being selective in the new scenes that we generate and add to the training set. This will help us achieve the best results with the least amount of training data added. The method selects scenes based on a “quality” criterion. To that end we investigate two criteria for evaluating the quality of new scenes namely, one based on entropy and another one known as the breaking-ties criterion. The initial results present promising capabilities of this solution for four RS scene datasets in enhancing the accuracy of classification.