Image-quality assessment for full-disk solar observations with generative adversarial networks^⋆

¹ University of Graz, Institute of Physics, Universitätsplatz 5, 8010 Graz, Austria
e-mail: robert.jarolim@uni-graz.at
² University of Graz, Kanzelhöhe Observatory for Solar and Environmental Research, Kanzelhöhe 19, 9521 Treffen am Ossiacher See, Austria
³ Skolkovo Institute of Science and Technology, Bolshoy Boulevard 30, Bld. 1, Moscow 121205, Russia

Received: 17 June 2020
Accepted: 24 August 2020

Abstract

Context. In recent decades, solar physics has entered the era of big data and the amount of data being constantly produced from ground- and space-based observatories can no longer be purely analyzed by human observers.

Aims. In order to assure a stable series of recorded images of sufficient quality for further scientific analysis, an objective image-quality measure is required. Especially when dealing with ground-based observations, which are subject to varying seeing conditions and clouds, the quality assessment has to take multiple effects into account and provide information about the affected regions. The automatic and robust identification of quality-degrading effects is critical for maximizing the scientific return from the observations and to allow for event detections in real time. In this study, we develop a deep-learning method that is suited to identify anomalies and provide an image-quality assessment of solar full-disk Hα filtergrams. The approach is based on the structural appearance and the true image distribution of high-quality observations.

Methods. We employ a neural network with an encoder–decoder architecture to perform an identity transformation of selected high-quality observations. The encoder network is used to achieve a compressed representation of the input data, which is reconstructed to the original by the decoder. We use adversarial training to recover truncated information based on the high-quality image distribution. When images of reduced quality are transformed, the reconstruction of unknown features (e.g., clouds, contrails, partial occultation) shows deviations from the original. This difference is used to quantify the quality of the observations and to identify the affected regions. In addition, we present an extension of this architecture that also uses low-quality samples in the training step. This approach takes characteristics of both quality domains into account, and improves the sensitivity for minor image-quality degradation.

Results. We apply our method to full-disk Hα filtergrams from the Kanzelhöhe Observatory recorded during 2012−2019 and demonstrate its capability to perform a reliable image-quality assessment for various atmospheric conditions and instrumental effects. Our quality metric achieves an accuracy of 98.5% in distinguishing observations with quality-degrading effects from clear observations and provides a continuous quality measure which is in good agreement with the human perception.

Conclusions. The developed method is capable of providing a reliable image-quality assessment in real time, without the requirement of reference observations. Our approach has the potential for further application to similar astrophysical observations and requires only coarse manual labeling of a small data set.