A Deep Learning Framework for Super-Resolution Reconstruction of SOHO/MDI Magnetograms

Reconstructing fine-scale magnetic field features from low-resolution historical solar magne-tograms, such as those from the Michelson Doppler Imager (MDI), is crucial for advancing solar physics and improving space weather prediction. This paper introduces RESM (Resolution Enhancement of Solar Magnetograms), a novel deep learning framework developed for super-resolution reconstruction of MDI magnetograms. RESM integrates Feature Enhancement Blocks (FEB) with a Convolutional Block Attention Module (CBAM) to enhance spatial detail while preserving structural magnetic features. Trained on 9,717 paired MDI-HMI magnetograms and validated on 1,332 pairs, RESM achieves a high correlation coefficient of 0.929 with HMI data, a PSNR of 55.6 dB, SSIM of 0.948, and a low RMSE of 0.071. These results significantly outperform conventional SR methods. The framework enhances the scientific utility of archival data and supports improved modeling and forecasting of solar flares. Future work will extend RESM to vector magnetogram reconstruction and evaluate cross-instrument generalization using Hinode/SP and Solar Orbiter PHI data.