DeepFilter. A machine learning technique for removing the hot AIA 304 Å channel component for the analysis of coronal rain

Abstract

The Atmosphere Imaging Assembly (AIA) 304 Å channel aboard the Solar Dynamic Observatory offers an unparalleled full-disk view of cool material at T ≈ 10 5 K emitted by the He II 304 Å spectral line. This opens the possibility for the in-depth and widespread analysis of the formation and evolution of small cool structures seen in the solar atmosphere. Of particular interest is the phenomenon of coronal rain, which has been linked to the overarching heating and cooling cycles of the solar corona. However, within the channel’s passband, hot diffuse emission from several ions is also included, leading to comparable intensity levels to the cool emission, particularly off-limb. This makes it very difficult to disentangle cool coronal rain from this hotter material. In this paper a novel morphological approach to separating these components called DeepFilter is investigated. This approach utilises a generative machine learning algorithm that can learn how to convert the AIA 304 Å images into the style of images obtained with the Interface Region Imaging Spectrograph (IRIS) 1400 Å, which has a similar temperature formation peak as for He II 304 Å but lacks this hot-component contamination. We find that the method produces good results, showing a clear reduction in the amount of hot-component material present in the final images while preserving the majority of the underlying cool structures. DeepFilter is compared to the recent physics-based RFit algorithm and is found to produce comparable results. Although the DeepFilter method is shown to perform worse at removing hot emission and material far from the limb, it performs comparably on other data – with the advantage of being far less data intensive – which makes it more effective for large-scale statistical analysis.