Problem

Coronal holes (CHs) are major sources of the fast solar wind, which is a significant component impacting space weather. These regions appear as distinct dark patches in extreme ultraviolet (EUV) images of the solar corona. A key challenge in solar physics is the accurate, real-time, and fully automated detection of these CHs. Existing methods often struggle to differentiate CHs from other dark features on the Sun and maintain consistent performance across various solar cycles, which exhibit significant changes in image contrast due to the presence of large-scale solar activity like active regions, solar flares, and coronal mass ejections. This variability makes relying solely on intensity-based thresholds problematic for reliable CH identification.

Method

The authors developed POP-CORN (Prevision Of Phenomena through Coronal-hole Outline Recognition with Neural-network), an automated tool designed to detect CH boundaries using a threshold-based technique powered by a neural network. Instead of directly processing EUV images, POP-CORN uses a neural network trained on 93 categorical and binary features. These features describe the properties and spatial distribution of large-scale solar corona structures, including active regions, solar flares (categorized by GOES class), coronal mass ejections, filaments, emerging flux, coronal dimming, and the phase of the solar cycle (minimum, rising, maximum, declining). Spatial distribution is represented by dividing the solar disk into quadrants. The model also incorporates sunspot classifications from the Mount Wilson magnetic classification. The neural network architecture includes hidden layers with ReLU activation, a sigmoid input layer, and a linear output layer that predicts a single optimal threshold value. To enhance model convergence and prevent overfitting, batch normalization and a dropout layer with a 0.3 rate are implemented. The model is optimized using the Adam optimizer, minimizing Mean Squared Error (MSE) and evaluated with Mean Absolute Error (MAE). Additionally, a separate linear neural network is used to handle conversions between threshold values for SDO/AIA 193Å and SoHO/EIT 195Å images, accounting for their differing pixel scales and sensitivities. The predicted threshold is then applied to binarized EUV images to delineate CH contours.

Results

POP-CORN demonstrated robust performance in detecting coronal hole contours across different phases of solar cycles 24 and 25. The model achieved a high Pearson correlation of 0.94 between predicted and validation threshold values, alongside a low Wasserstein distance of 0.11, indicating strong statistical agreement. Both qualitative and quantitative comparisons with ground truth (GT) observations showed excellent consistency, with the model successfully identifying CHs and avoiding misinterpretations of other dark features, even during periods of intense solar activity. Statistical validation using Hotelling's T-squared test largely confirmed that POP-CORN's contours were in agreement with the GT (p-value > 0.05 in most cases). Furthermore, POP-CORN qualitatively compared well with, and in some instances outperformed, other state-of-the-art CH detection tools like SPOCA, ACWE, CHARM, CHIMERA, and CHRONNOS, especially in capturing smaller CHs without overestimating filaments. However, for Solar Cycle 23, the model showed a tendency to overestimate CH regions, particularly during solar minimum and rising phases. This was attributed to the greater sensitivity of SoHO/EIT images to minor threshold variations and the limitations of the current SDO-SoHO threshold conversion, which the authors plan to refine in future work.

Implications

This research underscores the critical role of large-scale solar corona structures' properties and locations in accurate coronal hole detection. The findings suggest that manually incorporating these specific physical features into neural network training offers a more effective approach than training directly on raw EUV images, leading to more consistent and reliable CH identification. POP-CORN's ability to dynamically determine optimal threshold values, even in challenging conditions like the presence of numerous bright active regions or flares, is a significant advancement for solar observations. The tool is envisioned to be a crucial component in the WindTRUST project, where it will be integrated into a pipeline for validating solar wind models, such as the WindPredict polytropic (WP-γ) model. This integration will contribute to developing a fully automated system that provides quantitative scores for solar wind predictions, thereby enhancing space weather forecasting capabilities.