Convolutional Neural Networks for the Identification, Characterization and Modality Translation of Solar Active Regions from Ground-based Observations
(2025)
Files
NielsSayez_revised.pdf
Open Access - Adobe PDF
- 28.4 MB
Details
- Authors
- Supervisors
- Abstract
- This thesis, conducted in cooperation with the Royal Observatory of Belgium (ROB), presents novel deep learning approaches for monitoring solar active regions (AR) and enhancing solar observation data. Active regions, visible as dark spots named 'sunspots' on the solar surface, are key drivers of space weather phenomena. The first contribution of this thesis is the introduction of a fully automated framework for identifying, grouping, and classifying sunspots in white light imagery showing the solar surface, known as the photosphere. Using convolutional neural networks trained on automatically generated pseudolabels, this method effectively detects sunspots despite atmospheric disturbances inherent to ground-based observations. The system then successfully organizes individual sunspots into groups corresponding to ARs and classifies them according to the McIntosh system, matching the accuracy of ROB's manual catalog. The second contribution addresses gaps in solar observation data, particularly those from missing historical observation modes, by developing deep learning models that synthesize unavailable data types from existing ones. These models generate Calcium K-line images showing the low solar atmosphere, or chromosphere, from white light images, and magnetograms from extreme ultraviolet images. We show that modulating the internal representations of our models minimizes reconstruction errors and artifacts, increasing physical faithfulness. These advances demonstrated how deep learning can serve as a powerful tool to augment human expertise in solar physics. By automating routine tasks and extending observational capabilities, these methods enable researchers to focus on complex analysis and interpretation, ultimately contributing to more reliable space weather forecasting essential for protecting Earth's technological infrastructure from solar activity impacts.
- Affiliations
Citations
Sayez, N. (2025). Convolutional Neural Networks for the Identification, Characterization and Modality Translation of Solar Active Regions from Ground-based Observations. https://hdl.handle.net/2078.5/247551