A Bias-free Deep Learning Approach for Automated Sunspot Segmentation

Solar activities significantly influence space weather and the Earth's environment, necessitating accurate and efficient sunspot detection. This study explores deep learning methods to automate sunspot identification in solar satellite images, keeping personal bias to a minimum. Utilizing observations of the Solar Dynamics Observatory, we leverage active-region data from the Helioseismic Magnetic Imager active-region patches to locate sunspot groups detected between 2011 and 2023. The Morphological Active Contour Without Edges technique is applied to produce pseudo-labels, which are utilized to train the U-Net deep learning architecture, combining their strengths for robust segmentation. Evaluation metrics-including precision, recall, F1-score, intersection over union, and Dice coefficient-demonstrate the superior performance of U-Net. Our approach achieves a high Pearson correlation coefficient of 0.97 when compared with the sunspot area estimation of the Space Weather Prediction Center and 0.96 in comparison with the Debrecen Photoheliographic Data. This hybrid methodology provides a powerful tool for sunspot identification, offering the improved accuracy and efficiency crucial for space-weather prediction.