Classification of Solar Flare Magnitudes Using SDO/AIA Movies with 4D Convolutional Neural Networks
Abstract
Currently, solar flares are labeled with a magnitude based on a single global measurement of the sun -- the peak X-ray intensity over the flaring event, as measured by the GOES satellite. Given that solar flares are local events, human forecasters are then tasked with labeling the active region on the sun associated with the flare. This has potential for errors since it mandates using at least two separate observational systems. Interest has been growing in using imaging instruments to classify flare intensity and location simultaneously. In addition, imaging instruments such as the Solar Dynamics Observatory (SDO) Atmospheric Imaging Assembly (AIA), which provides full-Sun images in ultraviolet and extreme ultraviolet wavelengths, are increasingly used in machine learning (ML) solar flare prediction models, as these images may reveal more features associated with flaring than the photospheric magnetic field data that has been mostly used to date. We demonstrate the use of AIA image cutouts of solar active regions to characterize the peak X-ray magnitude of solar flares via ML regression to the GOES measurements, offering an alternative to using the GOES flare catalog for event location and data labelling. We use a 4D Convolutional Neural Network (CNN) algorithm trained on a temporal series of AIA images in various wavelengths, with the corresponding outputs being the GOES flare magnitude for the event. However, a challenge is that solar flare peak times and the length of flaring vary as a function of wavelength, e.g., the flare peak time in the SDO/AIA 171 Angstrom bandpass (as determined, e.g., by the maximum size of saturated pixels in the flare region) can be as much as 40 minutes after the flare peak in the GOES data). We will address this complication in constructing training databases for the ML algorithm and present preliminary results of our regression modeling.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFMNG45B0571V