Speaker
Description
CHESS expands the training of two baseline Convolutional Neural Networks (CNNs) to obtain a more efficient, least-biased CNN model for segmenting coronal holes (CHs).
Our two CNNs are based on (i) a U-Net and (ii) a Res-U-Net architecture. These two CNNs have been pre-trained with the coronal hole (CH) boundary data from the Heliophysics Events Knowledgebase (HEK). These initial, pre-training data of the CH boundaries are obtained by the Spatial Possibilistic Clustering Algorithm (SPoCA) applied to images of the Atmospheric Imaging Assembly (AIA) onboard the Solar Dynamics Observatory (SDO), using the extreme ultraviolet (EUV) 193-Å filter. In many instances, this algorithm cannot differentiate between a CH and another solar structure called "filament". Our project overcomes this limitation by adding ground-based observations of the He I 10830 Å spectral line, which is able to provide such disambiguation.
Full disk images in this chromospheric line primarily host three desirable traits, namely observability by more accessible ground telescopes, providing a clear distinction between filaments and CHs, and a lack of susceptibility to the blocking effect by the coronal emission from nearby Active Regions (ARs) in EUV observations. These full disk observations are provided by the National Solar Observatory (NSO) Vector Spectromagnetograph (VSM) instrument in the Synoptic Optical Long-term Investigations of the Sun (SOLIS) facility. The He I line is a prime candidate for extracting an alternate perspective on coronal hole boundaries and inclusion in a broader ensemble machine learning method in which boundary predictions from various base estimator routines are combined to a final estimator via stacking to yield a strong classifier of CHs. By using the He I imagery, we will further curate our CH training set that will be free of the contamination of filaments. This curated dataset is then used for improving the training of our two CNNs, which will be compared against our baseline versions that were contaminated by the presence of filaments. The best version(s) will then be made available to the community.
