Automatic sunspot detection through semantic and instance segmentation approaches

Abstract

The solar influence on space weather and terrestrial environment is substantial. Strong geomagnetic storm activity can significantly affect astronauts in orbit, communications and GPS systems and disrupt Earth’s power distribution networks, making continuous monitoring and forecasting of solar activity vital. Sunspots are magnetic disturbances in the photosphere characterized by their dark appearance in the solar disk, being directly related to phenomena that contribute to these intense storms, namely solar flares and coronal mass ejections. This article lies at the intersection between solar surveillance and computer vision by applying state-of-the-art deep learning algorithms in the automatic detection of sunspots and sunspot groups. Based on two techniques, semantic segmentation and instance segmentation, two algorithms are implemented to tackle both purposes, U-Net and Mask R-CNN respectively. The ground-truth dataset was built from the available Debrecen Heliographic Observatory (DHO) space-borne sunspot catalogues from 2010 to 2014. The best U-Net implemented model presented a 74.2% IoU, surpassing the detection results evidenced by the Automated Solar Activity Prediction System (ASAP). The instance segmentation approach, a novelty application technique for sunspot group detection and still a challenging task in computer vision, achieved 51.7 bounding box AP and 78.6% accuracy in predicting the number of sunspot groups.