Abstract:

Solar filaments are a regularly occurring feature of the solar atmosphere that provides crucial information on changes in solar activity and helps forecast solar weather. Most notably, filaments can give rise to coronal mass ejections (CMEs), a large expulsion of plasma and magnetic field from the Sun’s corona that can heavily impact Earth’s magnetosphere. However, filaments can be hard to detect across the entire solar disc, and methods that rely on human annotations, which are inherently costly and time-consuming, can lead to inconsistent mapping of solar phenomena. This project uses imagery from the Global Oscillation Network Group (GONG), which observes the full solar disk in the H-alpha band where filaments are most prominent. The first step uses preprocessing techniques to highlight filament features, along with the Segment Anything Model(SAM), to produce a first-pass filament segmentation. These SAM predictions are improved by incorporating physical constraints from known filament shapes, often connecting closely but separately located prediction masks. The second step trains a U-Net model on the machine-generated pseudo-labels to produce refined filament predictions. This model is validated against existing human-annotated filament mapping of the GONG H-alpha solar images. This self-training pipeline offers a scalable alternative to human annotations for filament mapping and the creation of a consistent, large-scale dataset. The dataset can serve as a new benchmark for solar filament detection models, and the self-training model can be adapted for automated analysis. The results are from the NOAA EPP CSC NERTO (in-residence at NOAA graduate internship) project conducted with NOAA mentor Rob Redmon of the NOAA National Centers for Environmental Information (NCEI). The NERTO aligns with NOAA CSC CESSRST-II’s goal to advance environmental data science and develop innovative remote sensing and machine learning capabilities that support NOAA’s mission. The NERTO project, A Self-Trained Deep-Learning Methodology for Automated Solar Filament Detection and Dataset Generation, also deepened the intern’s understanding of NOAA’s data stewardship practices, solar-terrestrial monitoring needs, and the application of artificial intelligence to large-scale environmental information systems.