Speaker
Description
Coronal Mass Ejections (CMEs) are critical drivers of space weather, requiring precise kinematic and morphological characterization to predict their geoeffectiveness. We previously demonstrated that fine-tuning the deep neural model Mask R-CNN on synthetic CME images, generated via Graduated Cylindrical Shell (GCS) shapes and raytracing, allows the automated segmentation of CME outer envelopes in coronograph images from SOHO/LASCO and STEREO/COR.
Following the validation of our core methodology, this work focuses on evaluating the model's performance when applied to solar observations from multiple instruments (not present in the training dataset). Specifically, we expand our validation framework by incorporating images from the Extreme Ultraviolet Imager (EUI/FSI) aboard Solar Orbiter and the ASPIICS coronagraph on Proba-3.
To quantify the network’s robustness, we assess standard segmentation metrics, including intersection over union, precision, and recall, and morphological parameters such as central position angle and angular width. All derived by comparing the predicted masks against a validation set of manually labeled images. This evaluation aims to establish the reliability of deep learning-based segmentation for future solar instruments and missions.