The next generation of astronomical surveys is set to generate unprecedented volumes of data. Yet, a fundamental trade-off still persists between sky coverage, image resolution, and multiband signal acquisition. In general, ground-based observatories offer wide coverage in multiple frequency bands but are limited in resolution by atmospheric blurring. On the other hand, space-based telescopes provide high-resolution imaging but over smaller areas of the sky, and can be limited by their size.
The primary goal of this thesis is to address the challenge of combining depth, wide sky coverage, and high spatial resolution by developing and validating advanced, deep learning-based deconvolution algorithms designed to enhance the resolution of ground-based images, thereby bridging the gap with space-based capabilities. Our investigation thoroughly explored different neural networks, ranging from foundational convolutional neural networks (CNNs) like U-Net to more advanced Vision Transformer (ViT) variants like the Swin Transformer UNet (SUNet). We performed our analyses, both visually and quantitatively, on simulated datasets based on the Cosmic Assembly Near-infrared Deep Extragalactic Legacy Survey (CANDELS). Eventually, the method's real-world performance was validated on data from the Very Large Telescope (VLT), Cosmic Evolution Survey (COSMOS), Gemini Observations of Galaxies in Rich Early ENvironments (GOGREEN), and VISTA (Visible and Infrared Survey Telescope for Astronomy).
Building on this study, we extend the deconvolution paradigm from single-band processing to a novel joint multiband deconvolution framework. This method is designed to combine relatively lower-resolution, multiband ground-based images with a single, higher-resolution space-based observation that has overlapping spectral coverage. The primary application of this framework is the joint analysis of data from the Vera C. Rubin observatory and the Euclid space telescope. We also apply this technique to a real-world scientific case, successfully deconvolving images of the Perseus cluster from the Canada-France-Hawaii Telescope (CFHT) by leveraging public data from the Euclid Early Release Observations (ERO). The results validate the framework's ability to recover fine-scale structures across a wide field of view.
This manuscript demonstrates that sophisticated, physically-motivated deep learning frameworks can greatly enhance the scientific value of ground-based astronomical data. By providing robust, fast, and generalizable tools for image deconvolution, this work enables detailed scientific analyses of galaxy structure and stellar populations previously only achieved with space-based observations.
