Fusion of CNN ensemble for Remote Sensing Scene Classification

Scene classification problem in remote sensing (RS) images has attracted many researchers recently. Different fusion methods have been widely used by the machine learning community to fuse classifiers. In this paper, a decision-level fusion method has been proposed to fuse a set of stat-of-the-art CNN classifiers, namely VGG-16, SqueezeNet, and DenseNet. First, the experiment proves that these classifiers do not make the same classification mistakes, i.e. most of the time at least one of them provides correct classification. Thus these three classifiers are diverse and hence complement each other. To exploit this discovery, a novel decision-level fusion method that combines the classification decisions using voting and confidence fusion techniques has been developed. To show the effectiveness of the proposed fusion method, the results demonstrate how the accuracy of the classification can be enhanced using fusion versus training individual networks. The preliminary results for the UC Merced dataset, the KSA multisensor dataset, Aerial Image Datasets (AID), Optimal31 dataset and Whurs19 dataset have been presented. Preliminary comparison to state-of-the-art methods show the promising capabilities of this solution and encourages to investigate this method further.