Toward Reliable Solar Flare Forecasting: Exploring Multi-Wavelength Data from the Solar Dynamics Observatory

Solar flares, originating from sudden energy releases in the Sun’s atmosphere, pose significant risks to spaceborne and terrestrial technological systems, including satellite operations, communications networks, and power grids. Accurate solar flare forecasting is therefore essential for mitigating these impacts and advancing space weather prediction capabilities. In this study, we present a comprehensive deep-learning-based approach utilizing multi-channel observations from the Solar Dynamics Observatory (SDO), a spaceborne remote sensing platform dedicated to solar monitoring. Our analysis focuses on classifying solar flares under three scenarios: C vs. 0, M vs. C, and M vs. 0, leveraging ten distinct image channels spanning photospheric magnetograms and extreme ultraviolet (EUV) wavelengths. We trained and evaluated three modern convolutional neural network architectures—ResNet50, GoogLeNet, and DenseNet121—using the True Skill Score (TSS) and Gini coefficient to assess performance. The results highlight the superior predictive power of magnetogram data, with additional contributions from EUV channels such as 94 and 211 Å. This work underscores the utility of combining multi-spectral solar observations with state-of-the-art deep learning architectures to capture subtle pre-flare signatures and improve flare prediction accuracy. Furthermore, the methodology and open dataset provide a reproducible benchmark for advancing solar flare forecasting, supporting the broader remote sensing and space weather research communities.