Speaker
Description
First Name: Diego
Last Name: Lloveras
Affiliation NASA Goddard Space Flight Center & George Mason University
All Authors: D.G. Lloveras, F. Iglesias, M. Sanchez, Y. Machuca, F. Manini, F. M. López , H. Cremades & A. Asensio Ramos
Abstract: Coronal Mass Ejections (CMEs) are a major driver of space weather, posing significant risks to technological infrastructure and societal systems. Accurate and timely identification of CMEs in coronagraph imagery is therefore essential for effective geoeffectiveness forecasting. In the last decade, deep convolutional neural networks (DCNNs) have achieved enormous improvements in solving various machine vision-related tasks. The absence of extensive, well-annotated datasets suitable for supervised training has hampered their application to CME segmentation. To address this data scarcity, we have produced a synthetic dataset of CME coronagraph images. This dataset is constructed by combining authentic quiet-Sun coronagraph observations with synthetic CMEs generated using the Graduated Cylindrical Shell (GCS) forward model, thereby capturing key morphological features. Our approach involves fine-tuning a pre-trained Mask R-CNN architecture to generate GCS-like masks of CMEs from individual differential coronagraphic images. This work presents our findings from the DCNN model trained explicitly for the identification and segmentation of CME outer envelopes. We analyze the performance of the DCNN in a test case dataset of images taken by Metis.