Automated Detection Techniques of Off-Limb Coronal Jets

This paper presents a unified framework for detecting and characterizing off-limb coronal jets in SDO/AIA 304 Å observations by combining a semi-automated rule-based algorithm, a deep-learning segmentation model, and a morphology-driven machine-learning classifier. First, a Semi-Automated Jet Identification Algorithm (SAJIA) is used to construct a decade-long, human-vetted catalogue of 1215 off-limb jets (longer than 14 Mm) during Solar Cycle 24, from which power-law energy–frequency distributions, butterfly-diagram–like latitudinal migration, quasi-annual oscillations, hemispheric asymmetries, and indications of active longitudes are derived. To address the identification limits of SAJIA, two machine-learning-based algorithms are developed in parallel. The first one, an Automated Jet Identification Algorithm (AJIA) based on a U-Net convolutional neural network, is trained on SAJIA-labelled true and false jets, achieving pixel-level segmentations with precision, recall, and true-negative rates of order 0.8, while remaining robust to the long-term degradation of the AIA detectors and enabling reliable high-cadence jet detection. The second one is a random-forest classifier trained on mathematical morphology descriptors of nearly 900000 off-limb structures, matched to SAJIA labels, and then applied to the full morphology catalogue with a conservative probability threshold, yielding 3452 new jet candidates, of which 3268 are confirmed by visual inspection as genuine jets that are predominantly faint, small, and concentrated at high latitudes early in the solar cycle. Together, these three methodologies provide a more complete and accurate census of off-limb coronal jets than any single approach alone, reveal that jets share scale-free energetics with flares and CMEs, and demonstrate how modern image processing and machine learning can be integrated to support future high-resolution, high-cadence studies of small-scale solar eruptive phenomena.