A Comparative Study on Geospatial Scene Reconstruction Using GANs and Diffusion Models

Cloud cover in satellite imagery significantly limits its use in environmental monitoring, land use analysis, and disaster management by obscuring critical surface details. Synthetic Aperture Radar (SAR) provides cloud-penetrating data, but its lower spatial resolution and radar backscatter characteristics limit its effectiveness in high-resolution optical imagery applications. This study explores the efficiency of state-of-the-art Generative Adversarial Networks (GANs) and Diffusion models in reconstructing cloud-covered regions in satellite images. Both models were trained and tested on identical datasets and evaluated qualitatively and quantitatively. The results show that Diffusion models slightly outperform GANs, although the difference in performance is minor. However, for large-scale and time-sensitive analyses, where speed and resource efficiency are critical, GAN models like Pix2Pix are a practical choice, particularly for images with thinner cloud cover. This study contributes to the growing field of geospatial data analysis, offering valuable insights into model selection for applications requiring cloud removal.