A Survey of Generative Adversarial Networks for Satellite Imagery: Applications, Image Types, Tasks, and Challenges
Hadeel Sameer Altahainah, Osamah Mohammed Alyasiri, Mohd Halim Mohd Noor · IEEE Access · 2025
Satellite images are crucial in diverse fields, including agriculture, urban planning, and environmental monitoring. However, traditional image processing methods are inadequate to address numerous challenges, like data-related, computational, and land cover-specific issues, which frequently impact the accuracy and quality of satellite imagery. Generative adversarial networks (GANs), a significant advancement in artificial intelligence (AI), have demonstrated potential for enhancing satellite image processing. Therefore, this research reviews various GAN variants. For instance, cycle-consistent GANs (CycleGANs), pixel-to-pixel image translation (Pix2Pix), deep convolutional GANs (DCGANs), Wasserstein GANs (WGANs), and conditional GANs (CGANs), with their related applications and categorization. Furthermore, state-of-the-art GANs show substantial enhancements in the satellite image areas. For example, the texture-enhanced self-attention GAN (TE-SAGAN) in texture improvement, the nadir-adjusted GAN (NaGAN) in off-nadir object recognition, and multispectral edge-filtered conditional GANs (MEcGANs) in cloud removal. The research presents five tasks of GANs in satellite imagery processing, such as data augmentation, segmentation, image reconstruction, translation, and surveillance. The most common evaluation metrics used to evaluate the performance of GAN-based satellite imaging systems are presented in this study. Finally, future studies should concentrate on improving these models for broad applications, prioritizing real-time data processing and overcoming computational challenges.