Speaker
Description
The study of space weather critically depends on the three-dimensional (3D) morphological and kinematic characterization of coronal mass ejections (CMEs). This process can be done via generic 3D point position estimation (using e.g., tie-pointing plus triangulation, differential emission measure tomography, polarization ratio or neural radiation fields techniques) or model-based 3D reconstructions. The former is typically limited by scarce spatial coverage and/or limited temporal resolution, critical for extended and dynamic events such as CMEs. Model-based approaches are usually done via a manual fit of a parametric geometric model to the CME outer shell observed in coronographic images, which is time-consuming and prone to subjective biases. To address these limitations, we present a deep learning framework structured into three fundamental stages. First, to mitigate the lack of pixel-level labeled observational data, we develop a methodology for generating large-scale synthetic training datasets; this is achieved by combining real coronal backgrounds with synthetic CME brightness images produced by applying ray-tracing techniques to a density distribution defined by the Graduated Cylindrical Shell (GCS) geometric model. Second, using these synthetic data, a Mask R-CNN convolutional network was trained in a supervised manner for the instance segmentation of the CME outer envelope in differential coronographic images. Its outcome demonstrates a high capacity to discriminate the CME from other radially moving structures without requiring explicit kinematic information. Finally the 3D reconstruction is addressed using a deep convolutional network architecture, designed to automatically infer the six parameters of the GCS model from input coronagraph images acquired from three different perspectives. Altogether, the implementation of these neural networks represents a significant advance toward the automated detection and 3D morphological characterization of CMEs, reducing human intervention and accelerating computational analysis.