A review and meta-analysis of generative adversarial networks and their applications in remote sensing

In International Journal of Applied Earth Observation and Geoinformation
Volume (Issue): 108
Peer-reviewed Article
cover image

Generative Adversarial Networks (GANs) are one of the most creative advances in Deep Learning (DL) in recent years. The Remote Sensing (RS) community has adopted GANs quickly, and reported successful use in a wide variety of applications. Given a sharp increase in research on GANs in the field of RS, there is a need for an in-depth review of the major technological/methodological advances and new applications. In this regard, we conducted a comprehensive review and meta-analysis of GAN-related RS papers, with the goals of familiarizing the RS community with the potential of GANs and helping researchers further explore RS applications of GANs by untangling challenges common in this field. Our review is based on 231 journal papers that were retrieved and selected through the Web of Science (WoS) database. We reviewed the theories, applications, and challenges of GANs, and highlighted the gaps to explore in future studies. Through the meta-analysis conducted in this study, we observed that image classification (especially urban mapping) has been the most popular application of GANs, potentially due to the wide availability of benchmark datasets. One the other hand, we found that relatively few studies have explored the potential of GANs for analyzing medium spatial-resolution multi-spectral images (e.g., Landsat or Sentinel-2), even though such images are often freely available and useful for a wide range of applications (e.g., urban expansion analysis, vegetation mapping, etc.). In spite of the applications of GANs for different RS processing tasks, there are still several gaps/questions in this field such as: 1) which GAN models/configurations are more suitable for different applications?) 2) to what degree can GANs replace real RS data in different applications? Such gaps/questions can be appropriately addressed by, for example, conducting experimental studies on evaluating different GAN models for various RS applications to provide better insights into how/which GAN models can be best deployed. The meta-analysis results presented in this study could be helpful for RS researchers to know the opportunities of using GANs and understand how GANs contribute to the current challenges in different RS applications.