Detection and Classification of Objects in Metis Solar Corona Images Using Deep Learning

This thesis presents an image-processing and machine-learning pipeline for the detection, anal- ysis, and classification of compact objects in Metis Level 0 visible-light images. The analyzed archive contained 42, 830 images acquired using different spatial configurations but the main analysis focused on 35, 781 images with dimensions of 1024 × 1024 pixels, as these represented the largest and most consistent subset. The initial detection procedure generated more than 6 million candidate objects. Preprocessing and filtering methods were applied to reduce detector artefacts, background variations, and isolated bright pixels. A subset of 7, 549 objects was then characterized using morphological and intensity features and analyzed through agglomerative clustering. The method identified five groups representing different compact, diffuse, extended, and elongated morphologies. The temporal behaviour of the detected candidates was also examined. A tracking algorithm grouped the detections into 5, 749 trajectories and separated them into one-appearance, station- ary, short-movement, and long-movement categories. Most tracks corresponded to candidates observed only once or at approximately fixed detector positions. Finally, a pretrained YOLO11s model was fine-tuned for the automatic detection of compact objects. On the independent test set, the model achieved a precision of 0.860, a recall of 0.856, an F1 score of 0.858, an mAP ∗ 50 of 0.930, and an mAP ∗ 50 : 95 of 0.642. The results demonstrate that the proposed pipeline can support the automatic analysis of com- pact sources in Metis L0-VL observations. The developed methods could also be adapted to other astronomical datasets containing faint, compact, moving, or transient features in noisy and non-uniform backgrounds.