A Comparison Study on Generative Adversarial Networks (GANs) for Satellite Image Upscaling
Brandon Kah Wei Li, Muhammed Basheer Jasser, Bayan Issa, Samuel-Soma M. Ajibade, Hui Na Chua, Richard T.K. Wong, Muzaffar Hamzah · 2024
Satellite images that are publicly available on the internet such as Google Maps are mostly low-quality and blurry which is not very useful for research or decision-making due to the lack of details. Recently image upscaling has been increasing in popularity to address several real-world problems where one approach is using Convolutional Neural Networks (CNNs). However, CNNs have difficulty in restoring fine details and this has led to the use of Generative adversarial networks (GANs) for image scaling. In this paper, three GAN models are trained which are Enhanced Super-Resolution Generative Adversarial Networks (Real-ESRGAN), Image Restoration Using Swin Transformer (SwinIR-Medium), and Dual Aggregation Transformer (DAT-2). The models are evaluated with well-known image metrics to determine each model’s effectiveness in upscaling low-resolution satellite images. The results show that all models can reconstruct the image to some degree where the best out of the three is DAT-2. Further fine-tuning of the best model by changing the loss functions has shown improvements in structural integrity for the upscaled image.