Speaker
Description
Solar active region detection from high-cadence EUV imagery is important for data-driven space weather monitoring, yet robust pixel-level characterization remains challenging because active and non-active bright structures can overlap in intensity and morphology. We present an interpretable machine-learning framework for solar active region detection in 193 Å images from the Atmospheric Imaging Assembly onboard NASA’s Solar Dynamics Observatory. The method converts the circular neighborhood around each sampled pixel into a one-dimensional data series using a 2D circular kernel transformation, enabling compact feature extraction from local image structure.
We investigate two feature-selection strategies: (i) statistical and entropy descriptors of the transformed series—median, 95th percentile, distribution entropy, and fuzzy entropy—and (ii) direct use of the transformed series values as input features. A support vector classifier with an RBF kernel is trained to assign each local region to one of three classes: quiet/no active region, non-eruptive bright zone surrounding an active region, and flaring active region. Using repeated stratified 10-fold cross-validation, the proposed framework yields classification accuracies of 0.900 with entropy features and 0.914 with statistical features, while the combined feature set reaches 0.940. Among individual descriptors, fuzzy entropy provides the best performance (A_KF = 0.895), outperforming distribution entropy (0.738), the 95th percentile (0.873), and the median (0.840).
The resulting classifications preserve a realistic spatial distribution of solar activity and support the definition of a generalized solar activity indicator derived from the fraction of flaring-region pixels. The proposed approach is scalable, interpretable, and suitable for automated event detection pipelines in heliophysics and space weather applications.