Semi-supervised center-based discriminative adversarial learning for cross-domain scene-level land-cover classification of aerial images

Abstract

Abstract Supervised scene classification of aerial images is of great importance in land-cover classification. However, annotating the labeled data required for the conventional classifiers and convolutional neural networks (CNNs), costs much manpower and time. Domain adaptation methods can overcome this problem, to some extent, by transferring previously labeled data to the new images, but the classification models trained from the previously labeled data are not discriminative enough for classifying aerial images from other domains because of the data distribution differences caused by the variations in sensors, natural environments, seasons, angles, locations, and so on. In order to solve this problem, we propose a semi-supervised center-based discriminative adversarial learning (SCDAL) framework integrating three parts, namely filtering out easy triplets, proposed hard triplet loss, and the adversarial learning with center loss. In the SCDAL framework, a difficulty measure is proposed to remove easy triplets under the constraint of between-class dissimilarity and intra-class similarity and better distinguish hard triplets. The filtered triplets are then used to train a more discriminative source feature extractor with the proposed hard triplet loss combining the hardest triplet loss and semi-hard triplet loss. Adversarial learning with center loss is also proposed to reduce the feature distribution bias between the source and target feature extractors and increase the discriminative ability of the target feature extractor. The SCDAL framework is tested on two large aerial images as a case study. The experimental results demonstrate that when adequate previously labeled data but limited labeled target data exist, the SCDAL framework is superior to most of the existing domain adaptation methods, with an improvement of at least 3% in overall accuracy. It is also proved that removing easy triplets, proposed hard triplet loss, and the adversarial learning with center loss all help to improve the overall accuracy.