CHESS: Coronal Hole Extraction with Semantic Segmentation
Description
This project aims at expanding the training of two Convolutional Neural Networks (CNNs) that we have already developed to obtain a more efficient, more accurate, and 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 for Coronal Hole Image Segmentation,with model (ii) currently being more accurate than model (i). 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 will overcome this limitation by adding ground-based observations of the He I 10830 Å spectral line, which is able to provide such disambiguation.
Detailed example
Binary masks of coronal holes and co-spatial maps of quantified uncertainties.
AI / analytics pattern
Computer Vision: AI that processes and interprets visual data (e.g., images and videos).
Automation level / stage
a) Pre-deployment – The use case is in a development or acquisition status.
Expected benefit
Improved segmentation of coronal holes with (i) higher accuracy, (ii) better performance, (iii) ready to be implemented in research-to-operations pipelines.
Controls / human review
ATO: Not reported; PIA: Not published