Machine Learning in Application to Automatic Noise Processing of Solar Spectrograms

Abstract

In this paper, a machine learning approach for processing solar spectral data were developed. Its performance was demonstrated using the example of the limb solar flare on July 17th, 1981. The results indicated that machine learning can be effectively utilized to transform between differently digitized spectra, fill gaps in unique experimental data, and undertake spectrum cleaning. Specifically, convolutional neural networks were devised to transform between reflective and transmissive scans of a solar flare spectrogram. This can be a convenient technique for treating the spectrograms of unique solar events, most notably increasing the proportion of observational spectra that can be further analyzed. Namely, in subsequent research, we will be able to confidently incorporate data such as that captured on the edges of spectrograms, which was previously deemed insufficiently reliable due to limitations of available processing techniques. This will consequently increase the number of spectral lines studied for certain observed events, which is paramount for constructing physical models, as the spectral peculiarities are expected to manifest consistently across different spectral lines. The developed approach also notably facilitates the detection and removal of impurities in the spectrograms. Previously, each distinct feature in the spectra was manually scrutinized to check its integrity in order to be excluded if identified as an impurity, such as a scratch or a dust particle. By employing the suggested protocol for treating the spectrogram, which includes scanning the spectrogram using multiple distinct techniques and then leveraging machine learning for comparison, the process of excluding impurities can now be automated. Furthermore, the spectrum areas affected by such exclusions can be restored, enabling further analysis.