부산대학교 도서관

The applicability of deep learning to remote sensing is rapidly increasing in accordance with the improvement in spatiotemporal resolution of satellite images. However, unlike satellite images acquired in near-real-time over wide areas, there are limited amount of labeled data used for model training. In this article, three kinds of deep learning applications—data augmentation, semisupervised classification, and domain-adapted architecture—were tested in an effort to overcome the limitation of insufficient labeled data. Among the diverse tasks that can be used for classification, rice paddy detection in South Korea was performed for its ability to fully utilize the advantages of deep learning and high spatiotemporal image resolution. In the process of designing each application, the domain knowledge of remote sensing and rice phenology was integrated. Then, all possible combinations of the three applications were examined and evaluated with pixel-based comparisons in various environments and city-level comparisons using national statistics. The results of this article indicated that all combinations of the applications can contribute to increase classification performance, even though the uncertainty involved in imitating or utilizing unlabeled data remains. As the effectiveness of the proposed applications was experimentally confirmed, enhancement in the applicability of deep learning was expected in various remote sensing areas. In particular, the proposed applications would be significant when they are applied to a wide range of study areas and high-resolution images, as they tend to require a large amount of learning data from diverse environments, owing to high intra-class heterogeneity.