Spatially variant neural blind deconvolution to reconstruct ground-based solar observations
Year: 2026
Type: article
Abstract: High-resolution ground-based solar observations provide access to the smallest spatial scales in the lower solar atmosphere. However, Earth's turbulent atmosphere imposes a significant limitation, by reducing the spatial resolution of such observations. To obtain high-quality observations, post facto reconstruction techniques need to be applied. Existing methods commonly rely on the assumption of isoplanatic conditions, where the distortions i.e., the point spread function (PSF), is assumed to remain spatially invariant across the field of view (FOV). However, this assumption limits the current reconstruction methods to small image patches. Extending image reconstruction methods to spatially varying PSFs is therefore essential for obtaining homogeneous high-resolution observations over large FOVs. We have developed a novel image-reconstruction method capable of recovering high-resolution solar observations in the presence of spatially varying atmospheric degradation. The method is designed to reconstruct large FOVs while simultaneously estimating pixel-wise varying PSFs from short-exposure image bursts, and it is applicable to various astrophysical observations. In this work, we present a physics-informed neural-network-based reconstruction framework for high-resolution solar imaging with spatially varying PSFs. The model represents the underlying object intensity distribution as a continuous function by mapping coordinate points $(x,y)$ directly to pixel intensities, o(x,y). In addition, a second neural network is introduced to estimate spatially varying PSFs by mapping coordinate points $(x,y)$ to pixel-wise PSFs, p_k(x,y), across the image. During training, synthetic short-exposure observations are generated by convolving the estimated object with the corresponding spatially varying PSFs directly in the image domain. The model parameters are optimized by minimizing the residuals between the synthesized and observed short-exposure image bursts, enabling the simultaneous recovery of both the object and the spatially varying PSFs. We evaluated the method using synthetic intensity images from radiative magnetohydrodynamic simulations degraded with spatially varying PSFs and noise to mimic realistic observational conditions. Quantitative analyses demonstrate that the model accurately reconstructs both the underlying high-resolution image and the spatially varying PSFs. We also applied the method to observations from the GREGOR and DKIST telescopes, as well as full-disk observations from Kanzelhöhe Observatory and compared the results with state-of-the-art reconstruction techniques. The reconstructions show improved spatial resolution, reduced noise, and enhanced recovery of fine-scale solar structures across large FOVs. We have made the code publicly available as an open-source package.
Source: Astronomy and Astrophysics
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,865
OpenAlex ID: W7223136217