Generative Adversarial Networks for High Resolution Satellite Imagery Reconstruction in Urban Planning

Abstract

Satellite imagery is used in the urban planning of land use, infrastructure monitoring, and disaster management. Nevertheless, the low-resolution satellite images do not provide the required level of detail in most cases and restrict their usefulness. The conventional techniques, such as interpolation, cannot factor in such important urban content as buildings, roads, and green areas. The current paper suggests that to resolve this problem, Generative Adversarial Networks (GANs) should be used to perform the super-resolution reconstruction of satellite images. An alternative GAN design consisting of a U-Net-based generator, residual blocks, and PatchGAN discriminator is proposed. The model employs adversarial loss, L1 loss, and perceptual loss (through VGG features) to boost the urban characteristics in high-resolution results. The usefulness of the suggested GAN model is measured based on a real-life dataset of  low-resolution images of Sentinel-2 and high-resolution images of Landsat, with emphasis on urban features in Coimbatore, India. Performance is evaluated through PSNR, SSIM, and urban-specific IoU, which are used to detect buildings. The results of the experiments indicate that the GAN model is better than traditional methods and other deep learning models. In particular, the model obtained a PSNR of 35.4 dB, SSIM of 0.94, and urban IoU of 0.81, which was much better than bicubic interpolation (PSNR = 28.2 dB, SSIM = 0.82, IoU = 0.65) and SRResNet (PSNR = 32.1 dB, SSIM = 0.89, IoU = 0.72). The GAN-generated high-resolution images are more detailed and accurate of the urban features and are useful in urban planning, zoning, and disaster risk management, which illustrates the potential of GANs to enhance the satellite imagery in urban planning reality.