High-Resolution Khalifa Satellite Imagery with AI-Powered Super Resolution Generative Adversarial Networks

Abstract

KhalifaSat is an Earth observation satellite equipped with multispectral remote sensing capabilities, allowing it to analyze and detect various features like buildings and roads. However, the satellite's limited resolution sometimes leads to blurry image details. To address this issue, super resolution techniques have been employed, and one notable approach is using Generative Adversarial Networks (GANs). This study introduces a novel algorithm called Hybrid SRGAN, which simplifies the Super-Resolution Generative Adversarial Network (SRGAN) model while maintaining high accuracy. The proposed modifications involve eliminating batch normalization layers and using bicubic interpolation for resizing low-resolution images. By removing batch normalization, the training speed improves, and the model generalizes better across different image types. The performance of the Hybrid SRGAN model is evaluated using metrics like SSIM, PSNR, and BRISQUE, demonstrating its superior reconstruction quality and overall performance, particularly in the luminance channel (YCbCr), when compared to state-of-the-art algorithms. This suggests that the Hybrid SRGAN algorithm offers a more efficient and effective solution for enhancing image resolution and quality from the satellite's data.