Enhancing image resolution of solar magnetograms: A latent diffusion model approach
Please login to view abstract download link
The spatial structure of the solar magnetic field plays a fundamental role in understanding solar dynamics and space weather. However, magnetograms from older instruments such as the Michelson Doppler Imager (MDI) suffer from limited spatial resolution, impeding detailed analysis of small-scale magnetic structures. Enhancing the resolution of legacy data is crucial for consistent studies across solar cycles and improved characterizations of active regions and solar flares. In this work, we introduce a novel latent diffusion model (LDM) approach to super-resolve MDI magnetograms to the higher spatial fidelity of the Helioseismic and Magnetic Imager (HMI). The LDM is trained using residuals on synthetically downsampled HMI data and fine-tuned on paired MDI-HMI images. Our method enhances MDI’s resolution from 2"/pixel to 0.5"/pixel. The super-resolved outputs are quantitatively assessed using peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), Fréchet inception distance (FID), and learned perceptual image patch similarity (LPIPS). We also evaluate whether critical physical features such as unsigned magnetic flux and active region size are preserved. Comparisons with baseline deterministic architectures and other diffusion-based models (e.g., DDPM variants) show that our LDM with residuals outperforms alternatives in both visual fidelity and physical consistency. A Fourier analysis confirms the model’s ability to resolve features below the original 2" threshold. Moreover, the probabilistic nature of LDMs enables us to estimate the reliability of the reconstructed features—something deterministic models cannot offer. Future research will explore temporal super-resolution to enhance the time resolution of MDI observations, enabling a more comprehensive understanding of solar dynamics in earlier cycles.