Automatic Characterization of Mid-latitude Multiple Ionospheric Plasma Structures from All-sky Airglow Images using Deep Learning Technique

This study presents a fully automated deep learning pipeline utilizing YOLOv8 and BoT-SORT to characterize the propagation parameters of mid-latitude ionospheric plasma structures from all-sky airglow images, demonstrating superior efficiency and reliability compared to previous semi-automatic methods for handling large datasets.

The Gist

Imagine the Earth's upper atmosphere, the ionosphere, as a giant, invisible ocean of charged particles (plasma) floating above us. Sometimes, waves and currents in this "ocean" create distinct patches or bands of plasma that move around. Scientists call these ionospheric plasma structures.

Why do we care? Because these moving patches can act like static on a radio or a glitch in a GPS signal, messing up our communication and navigation systems. To understand them, scientists use special cameras called All-Sky Airglow Imagers. These cameras take pictures of the faint, glowing light (airglow) that comes from the ionosphere, revealing these invisible plasma bands as dark or bright streaks in the sky.

The Problem: Too Much Data, Too Many Humans

For years, scientists have studied these glowing streaks by looking at the photos and manually drawing lines to measure how fast they are moving and which way they are going.

• The Analogy: Imagine trying to count every single car on a highway by watching a video feed and drawing a line on the screen for every car. It's slow, boring, and prone to human error. If you have 7 years of video footage (thousands of images), doing this by hand is impossible.
• The Limitation: Previous computer methods could only look at the "big picture" (like the average speed of traffic) but couldn't track individual cars (specific plasma bands) if there were many of them moving at once.

The Solution: A Digital "Eye" and a "Tracker"

This paper introduces a brand-new, fully automatic system that does the job of a human expert, but faster and without getting tired. They built a pipeline using Deep Learning (a type of AI that learns by example).

Here is how their system works, broken down into three simple steps:

1. The "Eye" (YOLOv8)

First, they taught a computer model called YOLOv8 (which stands for "You Only Look Once") to recognize these plasma bands.

• The Analogy: Think of YOLO as a super-fast security guard who can scan a crowded room and instantly point out every person, even if they are wearing different clothes or standing in the shadows.
• What it does: Instead of just saying "there is a band here," it draws a precise outline (a mask) around each individual band, separating them from the background.

2. The "Tracker" (BoT-SORT)

Once the "Eye" spots the bands, the system needs to know where they go in the next frame.

• The Analogy: Imagine the security guard spots 5 people walking. The BoT-SORT tracker is like a bouncer who gives each person a unique name tag (Track ID) and follows them through the crowd, ensuring they don't get mixed up with someone else.
• What it does: It follows each specific plasma band from one image to the next, even if the band changes shape or gets faint.

3. The "Three Judges" and the "Referee" (Quality Control)

Now that the system has tracked the bands, it needs to calculate their speed and direction. The authors didn't trust just one method, so they used three different mathematical techniques (Minima, MNCC, and Optical Flow) to do the math.

• The Analogy: Imagine three different judges in a talent show, each calculating the score for a performer using a different formula. Sometimes, one judge might be distracted or make a mistake.
• The Referee (Quality Filter): To make sure the final score is right, a "Referee" step checks the three judges' answers.
  • If all three judges agree closely, the system gives a Green Flag (High Confidence).
  • If they are a bit different, it gives a Yellow Flag (Use with caution).
  • If they are wildly different, it gives a Red Flag (Throw this data out).

Why This Matters

This new system is a game-changer for a few reasons:

1. It's Fast: It can process years of data in minutes, something that would take a human years to do.
2. It's Honest: The "Referee" step tells scientists exactly how reliable the data is. In the past, scientists had to guess if a measurement was good or bad; now, the computer gives them a quality score.
3. It Handles Crowds: It can track multiple plasma bands moving at the same time, even if they are interacting or crashing into each other, which previous methods struggled to do.

The Bottom Line

The authors have built a fully automated robot scientist that can watch the sky, spot the moving plasma bands, track them like a sports commentator, and calculate their speed and direction with a built-in "truth detector." This allows scientists to finally analyze massive amounts of data to understand the "weather" of our upper atmosphere, helping to protect our GPS and radio communications from future glitches.

Technical Summary

1. Problem Statement

The mid-latitude ionosphere hosts various plasma irregularities, such as Medium-Scale Traveling Ionospheric Disturbances (MSTIDs) and field-aligned plasma depletions. Understanding their electrodynamics requires accurate estimation of propagation parameters, specifically horizontal velocity and orientation (tilt angle).

• Limitations of Current Methods:
  • Semi-automatic approaches: Existing methods (e.g., Yadav et al., 2021b) rely on manual intervention to draw reference lines on airglow images. This introduces subjectivity, human error, and is time-consuming, making it impractical for large-scale, multi-year datasets.
  • Global Analysis Methods: Techniques like 3D spectral analysis or autocorrelation can process large datasets automatically but fail to distinguish and characterize individual plasma bands when multiple structures coexist in a single event.
  • Existing Deep Learning: While deep learning has been applied to equatorial plasma bubbles and single-band MSTIDs, there is a lack of frameworks capable of instance segmentation (distinguishing multiple overlapping structures) and continuous tracking of individual mid-latitude plasma structures from ground-based all-sky airglow imagery.

2. Methodology

The authors propose a fully automated pipeline consisting of three main stages: Instance Segmentation & Tracking, Multi-Method Parameter Estimation, and Quality Control.

A. Data Source

• Instrument: Multi-wavelength all-sky airglow imager at Hanle, India (32.77°N, 78.97°E).
• Wavelength: O(¹D) 630.0 nm (F-region emissions).
• Dataset: 1,768 preprocessed, unwarped grayscale images (511×511 pixels) spanning 2018–2025. Ground truth masks were generated using polygonal boundaries via VGG Image Annotator (VIA).

B. Step 1: Instance Segmentation and Tracking

• YOLOv8-seg: The authors utilized the You Only Look Once (v8) architecture with a segmentation head. It was chosen for its ability to perform single-pass regression, providing class labels, bounding boxes, confidence scores, and instance masks for each frame.
• BoT-SORT Tracker: Since YOLO processes frames independently, a BoT-SORT (Bag of Tricks - Simple Online Realtime Tracker) algorithm was integrated to assign unique Track IDs to each plasma structure across the image sequence.
  • Adaptation: Re-Identification (ReID) and Global Motion Compensation were disabled as the structures are non-rigid and the camera is stationary.
  • Mechanism: Uses Kalman Filters and Intersection over Union (IoU) to maintain track continuity, with relaxed detection thresholds to capture faint structures.

C. Step 2: Automatic Calculation of Propagative Parameters

Three distinct algorithms were applied to the segmented and tracked data to independently estimate horizontal velocity ($V$) and orientation:

1. Minima Method: Fits a line to the row-wise minimum intensity within the segmented mask. The perpendicular distance between minima lines of successive frames (same Track ID) yields velocity.
2. Maximum Normalized Cross-Correlation (MNCC): Uses template matching. A bounding box from the previous frame is correlated with the current frame to find the pixel shift corresponding to the peak correlation.
3. Optical Flow (Lucas-Kanade): Calculates motion vectors of pixels within the segmented masks. The bulk velocity is derived by averaging zonal and meridional components of these vectors.

D. Step 3: Quality Control (The "Quality Filter")

To ensure reliability, a logic-based filter combines the outputs of the three methods:

• Consistency Check: If the standard deviation of the three velocity estimates is $< 4$ m/s, the mean is accepted (Flag = 1).
• Outlier Rejection (Dixon's Q-Test): If the spread is larger, the method identifies and rejects the most extreme outlier.
  • Flag 1: High reliability (low variability).
  • Flag 0.5: Moderate reliability (outlier removed, moderate spread).
  • Flag 0: Poor quality (all estimates rejected or inconsistent).

3. Key Contributions

1. First Instance Segmentation for Mid-Latitude Structures: This is the first study to apply YOLO-based instance segmentation and continuous tracking to multiple coexisting mid-latitude plasma structures in all-sky airglow images.
2. Multi-Method Fusion: The integration of three independent physical algorithms (Minima, MNCC, Optical Flow) reduces reliance on a single estimation technique's biases.
3. Automated Quality Flagging: The introduction of a Quality Filter with a flagging system (0, 0.5, 1) allows researchers to automatically assess data reliability, a feature missing in previous semi-automatic methods.
4. Scalability: The pipeline eliminates manual intervention, making it feasible to process multi-year datasets (7+ years) for statistical studies, which was previously impractical.

4. Results

• Performance: The YOLOv8-seg model achieved high accuracy in segmenting and tracking individual bands, even in complex scenarios with multiple overlapping structures.
• Validation: The automated results were compared against a semi-automatic method (manual edge tracking) using a test set of 10 events (19 structures).
  • The automated velocities showed strong correlation with the semi-automatic method.
  • Approximately 26% of data points were filtered out by the quality control step, primarily due to rapidly evolving/decaying structures or electrodynamical interactions that caused high variance between the three algorithms.
• Case Studies:
  • 15 Sept 2018: The automatic method matched the semi-automatic results well.
  • 06 May 2019: In events with complex interactions and changing morphology, the three algorithms showed higher spread. The quality filter successfully flagged these as "moderate" or "poor" quality, preventing erroneous conclusions.
• Uncertainty: Estimated uncertainties were ~0.4° for tilt angle, ~10 m/s for Minima, ~11 m/s for MNCC, and ~8 m/s for Optical Flow.

5. Significance

This study represents a paradigm shift in ionospheric research from labor-intensive, subjective analysis to fully automated, reproducible, and high-throughput processing.

• Scientific Impact: It enables the characterization of individual plasma bands within complex, multi-structure events, providing deeper insights into electrodynamical interactions (merging, distortion, dissipation).
• Operational Utility: The framework is essential for handling the massive datasets generated by modern all-sky imagers and future satellite missions, allowing for long-term statistical studies of ionospheric dynamics without human bottlenecks.
• Data Integrity: The quality flagging system ensures that only reliable data is used for physical modeling, significantly improving the robustness of future ionospheric studies.