Learning Satellite Image Recovery Through Turbulence

This paper presents a study of deep learning approaches to image recovery using spatially-extended sequential observations of near-Earth satellites. Image recovery is often a prerequisite for use of ground-based extended imagery in space domain awareness (SDA) due to aberrations induced by atmospheric turbulence along the path from satellite to sensor. Traditional deconvolution-based image recovery methods are sensitive to factors such as observation sequence length and estimates of the point spread function (PSF), which has motivated recent interest in autoencoders and other learned approaches. However, no previous study has applied general state-of-the-art image restoration models to the space domain data. In this work, we evaluate the effectiveness of recent deep learning methods, specifically Generative Adversarial Networks (GANs) and Vision Transformers, for image restoration of satellites. We analyze the trade-offs between restoration quality, time, and computational complexity of each method. We experimentally demonstrate that deep learning models provide high-quality image restoration with less data than traditional deconvolution methods. We further optimize the most successful state-of-the-art model and demonstrate its efficacy in image restoration at a previously unseen degradation level (SNIIRS = 2.5). Our deep learning models are trained on simulated data from the SILO dataset and require no training on real data, yet they restore the most severely degraded real satellite imagery with state-of-the-art performance of 27.0 dB PSNR and 0.95 SSIM on the SILO dataset, as well as better visual results on the real satellite images.