Data Augmentation of Magnetograms for Solar Flare Prediction using
Generative Adversarial Networks
Abstract
Space weather forecasting remains a national priority in the United
States due to the impacts of events like solar flares to life on Earth.
High energy bursts of radiation originating from solar flares have the
potential to disrupt critical infrastructure systems, including the
power grid and GPS and radio communications. The rise of machine
learning and the development of higher-quality instruments has greatly
improved solar flare prediction models over the past decade. However,
the magnetogram data used for solar flare forecasting—taken by the
Solar and Heliospheric Observatory/Michelson Doppler Interferometer
(SOHO/MDI) and the NASA Solar Dynamic Observatory/Helioseismic and
Magnetic Imager (SDO/HMI) instruments—are incompatible due to
differences in the cadence, resolution, and size of the data.
Furthermore, many studies only focus on data from a single instrument
which disregards decades worth of potential training data that is
necessary to understand solar cycles. In this work, we show Generative
Adversarial Networks (GANs) can be used to super-resolve the historic
lower-quality SOHO/MDI data set to match SDO/HMI quality to create a
standardized magnetogram data set. The implementation of a Pix2Pix GAN
produced some undesirable artifacts in the synthetic image while image
translation methods CycleGAN and CUT preserved solar features present in
the data more accurately, even in the absence of paired data. The
resulting combined, higher-quality data set will be used to improve the
predictive power of current solar flare forecasting models.