🔭 astrophysics

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

This paper presents POP-CORN, an automated neural network-based tool that dynamically calculates detection thresholds for coronal holes in extreme ultraviolet images by incorporating large-scale solar structure properties, thereby enabling consistent and accurate identification across solar cycles 23, 24, and 25.

Original authors: K. H. P. Henadhira Arachchige, B. Perri, A. S. Brun

Published 2026-03-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: K. H. P. Henadhira Arachchige, B. Perri, A. S. Brun

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Finding the Sun's "Dark Spots"

Imagine the Sun as a giant, glowing campfire. Sometimes, there are huge, dark patches on this fire where the flames are actually weaker and cooler. In astronomy, we call these Coronal Holes (CHs).

Why do we care? Because these dark holes are like open windows in the Sun's atmosphere. They let out a super-fast stream of particles called the solar wind. When this wind hits Earth, it can mess up our satellites, GPS, and even cause beautiful auroras (Northern Lights). To predict space weather, we need to know exactly where these holes are and how big they are.

The Problem: The "Needle in a Haystack" (But the Haystack is Glowing)

For a long time, scientists have tried to find these holes automatically using computers. But it's tricky.

The Glare: The Sun is incredibly bright. Sometimes, other things on the Sun (like solar flares or dark filaments) look like holes, confusing the computer.
The Changing Light: The Sun goes through an 11-year cycle. Sometimes it's quiet, sometimes it's a chaotic storm of explosions. A computer program that works during a quiet day might get confused during a storm.
The "Real-Time" Gap: Most existing tools are either too slow to be useful right now, or they need a human to constantly tweak the settings.

The Solution: Meet POP-CORN

The authors of this paper created a new tool called POP-CORN (Prevision Of Phenomena through Coronal-hole Outline Recognition with Neural-network). Think of POP-CORN not as a camera that just "looks" at the Sun, but as a super-smart weather forecaster that uses a different kind of magic.

How It Works: The "Chef's Recipe" Analogy

Most computer vision tools try to learn by staring at millions of pictures of the Sun, hoping to figure out the pattern on their own. It's like trying to learn how to bake a cake just by looking at photos of cakes, without knowing the ingredients.

POP-CORN does it differently. Instead of just looking at the picture, it asks: "What is happening on the Sun right now?"

The Ingredients (The Inputs): The tool gathers a list of "ingredients" from a database. It asks:
- Is there a big storm (Active Region) nearby?
- Did a solar flare just go off?
- Is there a giant magnetic loop (Filament) that looks like a hole?
- Is the Sun currently in a "quiet" phase or a "stormy" phase of its 11-year cycle?
The Secret Sauce (The Neural Network): The tool uses a Neural Network (a type of AI) that acts like a master chef. It has been trained to know exactly how these "ingredients" change the brightness of the Sun's image.
- Example: If the AI knows a massive solar flare just happened nearby, it knows the whole image will look brighter. So, it automatically adjusts its "sensitivity" to find the dark holes, even if they are hidden in the glare.
The Output (The Threshold): Instead of drawing the holes itself, POP-CORN calculates the perfect "brightness cutoff" (a threshold). It tells the computer: "Anything darker than this specific number is a hole; anything brighter is just normal sun."

Why Is This Special?

The paper highlights three main wins for POP-CORN:

It's a Detective, Not Just a Camera: While other tools might mistake a dark filament (a long, thin rope of gas) for a hole, POP-CORN knows the difference because it "knows" the shape and location of filaments from its data. It's like a detective who knows that a shadow cast by a tree isn't a person.
It Handles the Chaos: When the Sun is having a tantrum (Solar Maximum) with lots of flares and bright spots, most tools get confused. POP-CORN stays calm because it accounts for the "noise" in its calculation.
It Works Across Time: The tool was tested on data from three different solar cycles (23, 24, and 25). It successfully adapted to the Sun's changing moods over decades.

The Results: A "Goldilocks" Performance

The team tested POP-CORN against human experts (who manually drew the holes on images).

During Quiet Times: It matched the humans almost perfectly.
During Stormy Times: It did better than most other automated tools, correctly identifying holes even when the Sun was covered in bright flares.
The One Hiccup: It had a little trouble with the oldest data (from the 1990s/early 2000s) because the cameras back then were a bit "blurrier" (lower resolution) than modern ones. The team is working on a "translator" to fix this.

The Future: The Automated Space Weather Pipeline

The authors aren't stopping here. Their ultimate goal is to plug POP-CORN into a larger machine that predicts the Solar Wind.

Imagine a fully automated factory line:

POP-CORN looks at the Sun and finds the holes.
It passes that info to a Wind Model (which simulates how the wind blows).
The system gives a score: "Here is our prediction for the solar wind hitting Earth in 3 days."

This creates a fully automatic system that helps us prepare for space weather without needing a human to stare at screens all day.

In a Nutshell

POP-CORN is a smart AI tool that doesn't just "see" the Sun; it understands the Sun's context. By feeding the AI information about solar storms, filaments, and the sun's current mood, it can automatically find the dark "windows" (coronal holes) that drive space weather, even when the Sun is being chaotic. It's a major step toward a fully automated system for protecting our technology from the Sun's temper.

1. Problem Statement

Coronal Holes (CHs) are regions of open magnetic field lines that serve as the primary source of fast solar wind, a critical component for space weather forecasting. While CHs appear as dark patches in Extreme Ultraviolet (EUV) images, their automated identification faces several challenges:

Ambiguity: Distinguishing CHs from other dark features (e.g., filaments, coronal dimming associated with Coronal Mass Ejections) is difficult using simple intensity thresholds.
Real-time Limitations: Existing models often lack real-time capability, are not fully automatic, or fail to separate CHs from bright features like active regions and solar flares.
Data Heterogeneity: Different observatories (SDO/AIA vs. SOHO/EIT) have different pixel scales and sensitivities, making a universal threshold difficult to apply across Solar Cycles 23, 24, and 25.
Training Limitations: Standard Convolutional Neural Networks (CNNs) trained directly on EUV images may struggle to capture the specific physics of large-scale structures or may be misled by image noise and artifacts.

2. Methodology

The authors developed POP-CORN (Prevision Of Phenomena through Coronal-hole Outline Recognition with Neural-network), a novel tool that predicts the optimal intensity threshold required to segment CHs in EUV images, rather than segmenting the images directly.

A. Data Sources

Imaging: LOS-EUV images from SDO/AIA (193Å, and composite 171+193+211Å) and SOHO/EIT (195Å) covering Solar Cycles 23, 24, and 25.
Contextual Data: Large-scale structure data retrieved from the Heliophysics Events Knowledge (HEK) database, including Active Regions (AR), Solar Flares (SF), Coronal Mass Ejections (CMEs), Filaments, and Coronal Dimming.
Solar Cycle Data: Sunspot numbers from SILSO and the specific phase of the solar cycle (Minimum, Rising, Maximum, Declining).

B. Feature Engineering

Instead of feeding raw images into the neural network, the authors engineered 93 distinct categorical (binary 0/1) features representing the state of the solar corona:

Spatial Distribution: Events are categorized by their location in solar quadrants.
Physical Properties: Solar flares are classified by intensity (X, M, C, A, B classes); Active Regions are classified by magnetic complexity (Mount Wilson classification).
Temporal Context: The specific phase of the solar cycle is included as an input variable.

C. Neural Network Architecture

Input: The 93 categorical features.
Output: A single continuous value representing the optimal intensity threshold for binarizing the EUV image to isolate CHs.
Structure: A feedforward neural network with hidden layers using ReLU activation, sigmoid for the input layer, and linear for the output.
Regularization: Batch normalization and a 0.3 dropout rate are used to prevent overfitting.
Optimization: Trained using the Adam optimizer to minimize Mean Squared Error (MSE).
Cross-Instrument Calibration: A secondary linear NN model was trained to map threshold values between SDO/AIA (high resolution, 0.6 arcsec/pixel) and SOHO/EIT (lower resolution, 2.6–5.26 arcsec/pixel) to ensure consistency across Solar Cycle 23.

D. Validation

Ground Truth (GT): CH contours manually identified by experts ("eye identification") on EUV images.
Statistical Tests:
- Pearson Correlation: 0.94 between predicted and true thresholds.
- Wasserstein Distance: 0.11, indicating small distributional differences.
- Hotelling's $T^2$ Test: Used to compare the spatial pixel coordinates of predicted vs. GT contours.

3. Key Contributions

Novel Input Strategy: Unlike traditional CNNs that process raw images, POP-CORN uses categorical features of large-scale structures as inputs. This explicitly encodes the physics of the solar corona (flare intensity, magnetic class, spatial distribution) into the model.
Threshold Prediction: The model outputs a dynamic threshold value rather than a static segmentation mask, allowing it to adapt to varying solar conditions and instrument characteristics.
Multi-Cycle Coverage: The tool is validated across Solar Cycles 23, 24, and 25, handling the transition from SOHO/EIT to SDO/AIA data.
Robustness to Noise: By incorporating flare and active region data, the model successfully distinguishes CHs from other dark features (like filaments) even during high-activity periods where brightness variations usually confuse threshold-based methods.

4. Results

Solar Cycle 25 (SDO/AIA): The model showed excellent agreement with Ground Truth (GT) across Minimum, Rising, and Maximum phases. Hotelling's $T^2$ tests confirmed statistical significance (fail to reject $H_0$ ) for all tested Carrington Rotations.
Solar Cycle 24 (SDO/AIA & SOHO/EIT): The model performed well across all phases, successfully identifying both large connected CHs and small isolated ones. The SDO-SOHO threshold conversion model proved effective.
Solar Cycle 23 (SOHO/EIT): The model performed well during Maximum and Declining phases. However, it showed overestimation during Minimum and Rising phases.
- Cause: The lower resolution of SOHO/EIT makes it more sensitive to small threshold variations, causing the model to include all dark regions as CHs when contrast is low.
Qualitative Comparison: Compared against state-of-the-art tools (SPOCA, ACWE, CHARM, CHIMERA, CHRONNOS), POP-CORN demonstrated superior ability to capture small CHs at mid-latitudes without misidentifying filaments, particularly in complex active regions (e.g., CR2170).

5. Significance and Future Work

Scientific Impact: The study confirms that manually incorporating the properties of large-scale structures (flare class, magnetic classification, spatial distribution) into the training data significantly improves CH detection accuracy compared to training solely on image pixels.
Operational Utility: POP-CORN provides a fully automated, real-time capable pipeline for CH detection, which is essential for space weather forecasting.
Future Integration: The authors plan to integrate POP-CORN into the WindTRUST project pipeline to validate the WindPredict polytropic (WP-γ) solar wind model. This will create a fully automated system providing quantitative scores for solar wind predictions based on observed CH boundaries.
Improvements: Future work focuses on refining the SDO-SOHO conversion model to eliminate overestimation in Solar Cycle 23 data.