A Robust Deep Learning Framework for Prominence Detection through Composite Feature Representations

Solar prominences are dynamic structures suspended within the solar corona and are a manifestation of solar activity. Their evolution includes eruptions linked to coronal mass ejections, making their detection critical for space weather monitoring and forecasting. The vast amount of high-cadence data provided by missions such as Solar Dynamics Observatory/Atmospheric Imaging Assembly motivates the application of deep learning frameworks capable of assimilating large-scale datasets. However, previous studies have reported poor model performance caused by contamination from hot coronal emission from the EUV He ii 304 Å channel. Using an existing dataset of labeled prominences, we find that trained YOLOv5 object detection models exhibit a strong bias toward the 304 Å color map, rather than physically meaningful prominence features. We develop a further two models comprising three-channel images constructed through an original dataset preprocessing pipeline: (i) full-disk grayscale, full-disk enhanced corona, and disk-removed; (ii) same as (i) with all disk-removed images. Our pipeline corrects instrument degradation to maintain more consistent feature representations across the solar cycle. The composite model (i) achieves a mAP@50 of 0.749 and a recall of 78% on the test set, outperforming previous bounding-box methods. Visual analysis of the composite models reveals that many apparent false positives are valid unlabeled prominences. We additionally demonstrate cross-instrument generalization by testing the composite model on SUVI image data. By examining dataset biases that propagate into model predictions, we provide recommendations for robust dataset construction. We present a reliable, physically motivated, and versatile deep learning model to automatically detect prominences in EUV images, providing a framework beneficial for space weather applications.