Identification of Coronal Holes on AIA/SDO Images Using Unsupervised Machine Learning

Abstract

Through its magnetic activity, the Sun governs the conditions in Earth’s vicinity, creating space weather events, which have drastic effects on our space- and ground-based technology. One of the most important solar magnetic features creating the space weather is the solar wind that originates from the coronal holes (CHs). The identification of the CHs on the Sun as one of the source regions of the solar wind is therefore crucial to achieve predictive capabilities. In this study, we used an unsupervised machine-learning method, k-means, to pixel-wise cluster the passband images of the Sun taken by the Atmospheric Imaging Assembly on the Solar Dynamics Observatory in 171, 193, and 211 Å in different combinations. Our results show that the pixel-wise k-means clustering together with systematic pre- and postprocessing steps provides compatible results with those from complex methods, such as convolutional neural networks. More importantly, our study shows that there is a need for a CH database where a consensus about the CH boundaries is reached by observers independently. This database then can be used as the “ground truth,” when using a supervised method or just to evaluate the goodness of the models.

Machine learning in solar physics

“…One of the earlier identification methods is the spatial possibilistic clustering algorithm 4 (SPoCA) which is implemented as part of JHelioviewer 5 (Barra et al 2008(Barra et al , 2009Verbeeck et al 2014). Other approaches are based on segmentation techniques together with ML algorithms (Reiss et al 2015), or on the fuzzy (Colak and Qahwaji 2013) and k-means clustering (Inceoglu et al 2022). The SPoCA method, based on an unsupervised fuzzy clustering method (Barra et al 2008), is a generalization of the k-means clustering discussed in Sect.…”

“…Recently, k-means has been implemented for the identification of CH by Inceoglu et al (2022). Three of the AIA/SDO wavelengths (171, 193 and 211 Å) were used in different combinations, individual channels, 2-channels (2CC) and 3-channels (3CC) composites, for building the data sets.…”

Efficient identification of pre-flare features in SDO/AIA images through use of spatial Fourier transforms

“…An even more surgical approach to flare prediction is suggested by the availability of high-spatial-resolution, multiwavelength, EUV observations obtained with the SDO/AIA telescope, which has provided (along with other data) full-disk 4, 096 × 4, 096 pixel images of the Sun, with 0′′.6 × 0′′.6 resolution, in seven different EUV wavelengths (93 Å, 131 Å, 171 Å, 193 Å, 211 Å, 304 Å, 335 Å) every 12 s since its launch in 2012; this represents an unprecedented archive resource for the investigation of the physical processes that lead to the onset of a solar flare. Due to the immense amount of data available, and an open policy for data distribution, this archive has attracted a lot of attention in the past few years for the development of machine learning methods (e.g., Huang et al, 2018;Galvez et al, 2019;Inceoglu et al, 2022). Although several recent works have dealt with combining HMI and AIA data for improving the predicting capabilities of machine learning techniques (e.g., Jonas et al, 2018;Nishizuka et al, 2018), to the best of our knowledge, no attempt has been made to forecast solar flares by using information encoded in the AIA data alone.…”

Forecasting solar flares with a transformer network

“…Various AI methods are used in heliophysics to tackle problems ranging from transportation of the charged particles throughout the heliosphere (Inceoglu et al, 2022a) to detecting magnetic activity structures on the Sun (Jarolim et al, 2021;Inceoglu et al, 2022b) and to predicting CMEs (Inceoglu et al, 2018;Raju and Das, 2023), as well as predicting the occurrences of solar flares with a forecast window of 24-48 h. In 2015, Bobra and Couvidat (2015) trained a support vector machine (SVM) algorithm on SHARPs data to classify whether an AR would emit a flare. In 2018, Nishizuka et al (2018) used vector magnetograms, along with 1,600 and 131 Å filter images and soft X-ray emission light curves, to train a deep neural network to predict whether an AR would release a flare of a specific magnitude within 24 h. The authors later applied this algorithm to develop an operational solar flare prediction model (Nishizuka et al, 2021).…”