POP-CORN: Validation of a new coronal hole detection tool based on neural networks

TL;DR

This paper introduces POP-CORN, a neural network-based tool for automatic, real-time detection of coronal holes in solar EUV images, improving accuracy by considering large-scale solar structures.

Contribution

The work presents a novel AI-driven method that automatically determines detection thresholds for coronal holes across multiple solar cycles, incorporating large-scale solar features.

Findings

Enhanced detection accuracy by including large-scale structure properties.

Automatic threshold determination enables real-time coronal hole identification.

Consistent detection across different solar activity periods.

Abstract

The properties and spatial distribution of large-scale structures of the solar corona determine the observed solar wind structure at 1 au. Coronal holes are a major source of fast solar wind, an important geo-effective component, and appear as large dark patches in extreme ultraviolet images. Solar observatories provide images of the solar corona at different wavelengths, enabling identification of coronal hole morphology and other large-scale structures along a given line of sight. The problem is that although models exist, few work in real time, separate coronal holes from other dark features, or are fully automatic and suitable for comparison with models. The main goal of this work is to develop an automatic threshold-based coronal hole detection tool across solar cycles 23, 24, and 25 using artificial intelligence. The only user input is the date, enabling retrieval of the threshold value used to detect coronal hole contours in line-of-sight extreme ultraviolet images from SDO/AIA and SoHO/EIT. We retrieve data affecting the threshold due to contrast changes from the Heliophysics Events Knowledge database for large-scale features such as active regions, solar flares, coronal mass ejections, and filaments, and engineer them to train the neural network model (POP-CORN). The model input comprises categorical features of large-scale structures in the solar corona, including spatial distribution and properties such as solar flare class by intensity. The neural network model is trained to achieve higher accuracy and determines the threshold needed to detect coronal holes, allowing their boundaries to be identified automatically and consistently. We conclude that properties of large-scale structures affect the determination of coronal hole regions, and incorporating these properties into training improves detection.