🔭 astrophysics

Automatic detection of solar filament oscillations I: Multi-scale spectral pipeline

This paper presents an automatic, multi-scale spectral pipeline that combines deep learning and statistical analysis to detect solar filament oscillations in GONG H-alpha data, significantly increasing detection sensitivity and scalability compared to traditional manual methods.

Original authors: Guillem Castelló, Manuel Luna, Jaume Terradas

Published 2026-07-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Guillem Castelló, Manuel Luna, Jaume Terradas

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

Imagine the Sun as a giant, bubbling pot of plasma. Sometimes, huge, dark ribbons of cooler gas, called filaments, float above the surface like clouds in the sky. These filaments aren't static; they wiggle, sway, and oscillate, much like a guitar string being plucked. By studying how they wiggle, scientists can learn about the invisible magnetic "strings" holding them up.

For a long time, finding these wiggles was like trying to find a specific needle in a haystack by looking at the haystack with a magnifying glass, one tiny square inch at a time. Scientists had to manually draw lines (slits) across the sun's image, watch the video, and guess where the interesting movement was happening. This was slow, exhausting, and often missed the quieter, smaller wiggles because the human eye gets tired and biased toward the loudest, most obvious movements.

The New "Smart Camera" Approach

The authors of this paper built an automatic pipeline—a smart computer program designed to do the heavy lifting. Think of it as a new kind of camera lens that doesn't just look at the sun, but looks at it in a very specific, clever way.

Here is how their method works, using some everyday analogies:

1. The "Blurry Photo" Trick (Multi-Scale Analysis)
If you look at a crowd of people from very close up, you see individual faces, hair, and tiny movements. It's chaotic and hard to see the big picture. If you step back and squint (blur the image), you stop seeing the individuals and start seeing the crowd moving as a whole.

The researchers' program does exactly this. It takes the sharp, high-definition video of the sun and creates many "blurry" versions of it.

Small blur: Looks at small patches. This is noisy and full of random static.
Medium blur: Smooths out the noise.
Large blur: Looks at the whole filament as one big unit.

The program checks all these different "blur levels" at once. If a wiggle is real and affects the whole filament, it will show up clearly in the medium and large blurry versions. If it's just a random glitch in one tiny spot, it disappears when you blur the image. This helps the computer ignore the "static" and find the "music."

2. The "Smart Filter" (Deep Learning & Calibration)
The program uses a "brain" (a neural network) that has been trained to know what normal, random sun noise looks like. It's like a security guard who knows the difference between a harmless bird flying by and a real intruder.

However, the sun's "weather" changes every day (clouds, atmospheric turbulence). A fixed rule wouldn't work. So, the program uses a special calibration trick (called conformal prediction) to adjust its sensitivity every single day. It asks, "Given today's specific sky conditions, how loud does a wiggle need to be to count as real?" This ensures the computer doesn't get fooled by bad weather or random noise.

3. The "Group Hug" (Cross-Scale Consistency)
The most important rule the program follows is: "If it's real, it must be seen from multiple distances."

If the program finds a wiggle in a small blurry patch, but not in the medium or large ones, it says, "Nah, that's probably just a glitch." It only keeps the wiggles that appear consistently across at least four different "blur levels." This is like a detective only arresting a suspect if four different witnesses, standing at different distances, all agree on what they saw.

What Did They Find?

The team tested their new system on data from the first two weeks of January 2014.

The Old Way: A previous manual study found 22 wiggles in that time.
The New Way: Their automatic pipeline found 91 wiggles.

That is more than four times as many!

They found that filaments are wiggling much more often than we thought. Some days, the manual study said "nothing happened," but the new pipeline found six or more distinct wiggles. They even found a specific wiggle on January 13th that no one had ever reported. To prove it was real, they went back and manually checked that specific spot with the old-school method, and sure enough, the wiggle was there, just too subtle for the human eye to catch in a quick glance.

The Bottom Line

This paper doesn't just find more wiggles; it changes how we look at the sun. It moves us from a slow, manual process where we only see the "loud" events, to a fast, automatic process that catches the quiet, subtle, and constant dance of the sun's magnetic ribbons.

The authors say this is just the first step. Now that they have a reliable, automated way to find these events, they can build a massive catalog of sun wiggles to study how the sun behaves over years and even decades. They aren't claiming to cure diseases or predict the weather on Earth yet; they are simply saying, "We finally have a way to see the sun's heartbeat clearly and consistently."

Technical Summary: Automatic Detection of Solar Filament Oscillations I: Multi-scale Spectral Pipeline

Problem Statement
Solar filament oscillations serve as critical probes for understanding the magnetic structure and stability of prominences. While long-duration, high-cadence H $\alpha$ archives from the Global Oscillation Network Group (GONG) offer the potential for systematic studies, current detection methods remain largely manual. These traditional approaches rely on visual inspection, manually placed slits, and time–distance diagrams. Such methods are labor-intensive, difficult to scale to large datasets, and inherently biased toward the most conspicuous, large-amplitude events. Furthermore, pixel-wise spectral analyses often fail to distinguish global filament oscillations from local intensity fluctuations, resulting in fragmented spatial signals that lack coherence. Previous attempts at automation (e.g., Castelló et al. 2025) utilized fixed confidence thresholds that could not adapt to the variable quality of ground-based data.

Methodology
The authors present a novel, automatic pipeline designed to detect spatially coherent oscillations in GONG H $\alpha$ image sequences. The framework integrates deep learning for segmentation with a multi-scale spectral analysis strategy. The workflow proceeds as follows:

Data Preprocessing: Daily GONG H $\alpha$ sequences are coaligned, corrected for limb darkening, and normalized to remove large-scale intensity gradients using Zernike polynomials. Off-disk pixels are masked to restrict analysis to on-disk filaments.
Filament Detection: Filaments are identified using deep-learning models (specifically a DETR-based detector and a U-Net segmenter from Reche & Cid 2024) to generate bounding boxes and binary masks.
Multi-scale Spectral Analysis: For each filament region, the image sequence is spatially averaged over overlapping kernels of varying sizes ( $N$ ), ranging from $N=71$ to $N=9$ . This multi-scale approach aims to suppress local noise while preserving collective filament motion.
Spectral Estimation: Lomb–Scargle periodograms are computed for the time series of each degraded pixel. The spectral background is modeled using a pre-trained one-dimensional Convolutional Neural Network (CNN) to estimate red-noise parameters ( $a f^{-\alpha} + b$ ).
Conformal Calibration: To address the rigidity of fixed thresholds, the pipeline employs a conformal prediction-inspired empirical calibration. It calculates residuals between observed spectra and the CNN-predicted background for non-filament regions, deriving a day-dependent multiplicative correction to establish significance thresholds (approximating a $5\sigma$ level).
Event Consolidation: Candidate detections are grouped by period and spatial location. A detection is retained in the final catalog only if it is supported by at least four distinct spatial scales, ensuring the signal is spatially coherent and not a localized artifact.

Key Contributions

Multi-scale Strategy: The introduction of spatial averaging over multiple kernel sizes to isolate coherent filament-scale oscillations from local fluctuations, addressing a primary limitation of pixel-wise analysis.
Adaptive Significance Calibration: Replacing fixed confidence levels with an empirical, day-dependent calibration derived from conformal prediction principles, allowing the pipeline to adapt to varying data quality.
Automated Pipeline: A fully reproducible framework that combines automated segmentation, CNN-based background estimation, and strict cross-scale consistency criteria to generate a robust catalog of oscillations.

Results
The pipeline was validated against the first two weeks of January 2014 GONG data, a period covered by the manual catalog of Luna et al. (2018).

Recovery of Known Events: The pipeline successfully recovered six events from the Luna et al. (2018) catalog, including the prominent 1 January 2014 oscillation with a period of $\sim76$ minutes.
Increased Detection Sensitivity: While the manual catalog reported 22 events (after duplicate removal) for this period, the automated pipeline identified 91 oscillatory events. This represents a more than four-fold increase in detection rate, averaging 6.5 events per day compared to 1.5 in the manual survey.
Discovery of New Events: The method identified significant activity on days where no events were previously reported (e.g., 2, 3, 13, and 14 January). A specific case on 13 January 2014 (period $\sim86$ min) was independently confirmed using a conventional time–distance diagram, verifying that the pipeline detects genuine oscillations often missed by visual inspection.
Period Distribution: Detected events exhibited periods ranging from approximately 20 to 126 minutes.

Significance and Claims
The paper claims that this framework provides a reproducible and scalable alternative to manual slit-based searches. By prioritizing spatial coherence through multi-scale analysis and calibrated significance thresholds, the method substantially increases detection sensitivity and reduces observational bias. The authors assert that filament oscillations are far more common than previously estimated, representing a near-constant dynamical signature of these structures.

The work establishes the foundation for future statistical studies of filament oscillations over extended GONG time intervals and across the solar cycle. The authors note that while the current implementation focuses on identifying location and period, it is not designed to track instantaneous frequency variations (e.g., wave packets), which would require time-frequency methods. The pipeline is presented as a robust tool for the automated discovery of dominant periodicities, paving the way for large-scale investigations into occurrence rates and triggering mechanisms in a forthcoming study (Paper II).

Technical Summary: Automatic Detection of Solar Filament Oscillations I: Multi-scale Spectral Pipeline

More like this