Fast Bayesian spectral analysis using convolutional neural networks: Applications to GONG Hα solar data

Abstract

Context. Solar filament oscillations have been observed for many years, but recent advances in telescope capabilities now enable a daily monitoring of these periodic motions. This offers valuable insights into the structure of filaments. A systematic study of filament oscillations over the solar cycle can shed light on the evolution of the prominences. Only manual techniques were used so far to analyze these oscillations.Aims. This work serves as a proof of concept and demonstrates the effectiveness of convolutional neural networks (CNNs). These networks automatically detect filament oscillations by applying a power-spectrum analysis to Hα data from the GONG telescope network.Methods. The proposed technique studies periodic fluctuations in every pixel of the Hα data cubes. Using the Lomb-Scargle periodogram, we computed the power spectral density (PSD) of the dataset. The background noise fits a combination of red and white noise well. Using Bayesian statistics and Markov chain Monte Carlo (MCMC) algorithms, we fit the spectra and determined the confidence threshold of a given percentage to search for real oscillations. We built two CNN models to obtain the same results as with the MCMC approach.Results. We applied the CNN models to some observations reported in the literature to prove its reliability in detecting the same events as the classical methods. A day with events that were not previously reported was studied to determine the model capabilities beyond a controlled dataset that we can check with previous reports.Conclusions. CNNs prove to be a useful tool for studying solar filament oscillations using spectral techniques. The computing times are significantly reduced for results that are similar enough to the classical methods. This is a relevant step toward the automatic detection of filament oscillations.