Identification and 3D morphological modeling of solar coronal mass ejections using deep neural networks

SANCHEZ, MARIANO; IGLESIAS, FRANCISCO; CISTERNA, FLORENCIA; MACHUCA, YAZMIN; LLOVERAS, DIEGO G.; MANINI, FRANCO; LOPEZ, FERNANDO; CREMADES, HEBE

San Juan

Congreso; 65a Reunion Anual de la Asociación Argentina de Astronomı́a; 2023

Asociación Argentina de Astronomía

Coronal mass ejections (CMEs), which are among the most magnificent solar eruptions, are a major driver of space weather and thus can have major negative technological and social implications. Given our current inability to forecast the occurrence of a CME, it is crucial to asses their geoeffectiveness once they are ejected. Particularly relevant for this task, are the identificcation and correct assesment of the CME 3D morphology in coronagraph images, routinely acquired from dedicated space solar observatories. In the last decade, Deep Neural Networks (DNN) have experienced enormous improvements in solving various machine vision-related tasks, particularly excelling at image recognition and segmentation. Recently, multiple DNN-based, machine vision models trained on very large datasets for image feature extraction, such as the Residual Learning Network (ResNet), have been made public to be use by the community as backbones for other application-specific developments. One issue when trying to use these deep models for CME segmentation or related tasks using coronagraph images, is that no large curated dataset exist in the literature that can be used for supervised training. To mitigate this, we produced a synthetic dataset of CME coronagraph images that incorporates the main features of interest, by combining actual quiet (no CME) coronagraph images with synthetic CMEs, the latter simulated using the Graduated Cylindrical Shell geometric model (GCS). Given that the exact 3D shape of the synthetic CME is known, this dataset can be used to do supervised training of a DNN-based model for different tasks. The trained model can then be used on actual CME coronagraph images, with its performance heavily depending on the quality of the synthetic dataset, i.e., if the relevant data features for the target task are included there. In this work, we present preliminary results of two DNN-based models. The first model is used to identify and segment the outer envelope of CMEs in a single image. This is done by fine tuning the pre-trained MaskR-CNN model, to produce a GCS-like mask of the CME present in a single differential coronagraph image. The second model, estimates the simplified 3D structure of the CME outer envelope from 2 and/or 3 simultaneous differential coronagraph images, acquired from different vantage points (spacecrafts). This model is implemented by adding a fully-connected, linear head to a pre-trained ResNet backbone, and is trained to produce the GCS model parameters that best fit the CME outer envelope in the input images.