🔭 astrophysics

StreakMind: AI detection and analysis of satellite streaks in astronomical images with automated database integration

StreakMind is an automated pipeline that utilizes a YOLO OBB model trained on a hybrid dataset to detect, characterize, and cross-identify satellite and Near-Earth Object streaks in astronomical images with high precision and recall, integrating results into a structured database to support large-scale surveys and space situational awareness.

Original authors: Rafael Carrillo Navarro, René Duffard, Pablo García-Martín, Javier Romero, Nicolás Morales, Luis Gonçalves

Published 2026-05-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rafael Carrillo Navarro, René Duffard, Pablo García-Martín, Javier Romero, Nicolás Morales, Luis Gonçalves

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 night sky as a giant, high-resolution movie screen. For decades, astronomers have been watching this screen to find tiny, moving dots that are asteroids or comets. But lately, the screen has become cluttered with "static"—bright, fast-moving lines left behind by thousands of human-made satellites and space junk. These lines, called streaks, are like pen marks drawn across a photograph, ruining the view and making it hard to see the real stars and asteroids.

Until now, finding these streaks was like trying to find a needle in a haystack by looking at every single piece of hay with a magnifying glass. It was too slow, too tiring, and impossible to do for the massive amounts of data modern telescopes collect.

Enter "StreakMind."

Think of StreakMind as a super-smart, tireless robot detective designed to scan these astronomical photos. Here is how it works, broken down into simple steps:

1. The Training: Teaching the Robot to See

The researchers didn't just show the robot real photos; they had to teach it what a "streak" looks like.

Real Data: They used thousands of real photos taken from the La Sagra Observatory in Spain.
Fake Data: Because real photos didn't have enough examples of very long streaks, they used computer code to "paint" fake streaks onto the images. It's like a teacher showing a student a few real examples of a dog, but then drawing hundreds of pictures of dogs of all shapes and sizes to make sure the student recognizes any dog, not just the ones they've seen before.
The Brain: They used a type of AI called YOLO (which stands for "You Only Look Once"). Imagine a security guard who can glance at a crowded room and instantly spot a specific type of person without needing to look at every face one by one. This AI was trained to spot the linear "pen marks" of satellites.

2. The Hunt: Finding the Streaks

Once trained, StreakMind looked at new photos.

The Filter: Sometimes, bright stars in the photo have "spikes" (like the starburst effect in photos). The AI might mistake these spikes for satellite streaks. StreakMind has a built-in filter that checks a cosmic phonebook (the Gaia star catalog) to see if a streak is actually just a star's glare. If it is, it ignores it.
The Stretch: The AI's first guess at where a streak starts and ends is often a little too short, like a rubber band that hasn't been pulled tight yet. StreakMind then performs a "photometric elongation"—it looks at the brightness of the streak and stretches the box around it until it captures the full, faint tail of the satellite.

3. The Detective Work: Connecting the Dots

Satellites move fast. If you take a series of photos a few seconds apart, the satellite will appear in a slightly different spot in each one.

The Chain: StreakMind acts like a detective connecting clues. It takes a streak from Photo A, calculates where the satellite should be in Photo B based on its speed and direction, and checks if there's a matching streak there.
The Tolerance: If the satellite is hidden behind a cloud for a moment (missing a photo), the system is smart enough to say, "Okay, it's still the same object," and keep the chain going, rather than giving up.

4. The ID Check: Who is it?

Once the system has tracked a streak across several photos, it asks: "Is this a known satellite?"

It compares the path of the streak against a database of known satellite orbits (ephemerides).
It uses a confidence score (like a weather forecast probability). If the match is very close, the confidence is high. If the match is a bit fuzzy, the confidence is lower. This helps astronomers know which identifications are solid and which need a second look.

5. The Filing System

Finally, all this information isn't just left in a pile. StreakMind automatically organizes everything into a neat, digital database. It creates a standard report (similar to how astronomers report finding new asteroids) that includes exactly where the streak was, how long it was, and which satellite it likely belongs to.

The Results

When the researchers tested this system:

It was incredibly accurate, finding 97% of the real streaks (Recall) and being right 94% of the time when it said it found one (Precision).
It could spot faint streaks that human eyes would likely miss because they were too tired or the streaks were too dim.
It did all this in a fraction of the time it would take a human to do the same job.

In short: StreakMind is an automated, AI-powered tool that cleans up the "noise" of satellite streaks from astronomical photos, identifies what those streaks are, and files the data away, freeing up human astronomers to focus on the actual science of the universe rather than cleaning up space junk.

Technical Summary: StreakMind – AI Detection and Analysis of Satellite Streaks

Problem Statement
The advent of wide-field, high-cadence astronomical surveys has generated data volumes that render manual inspection of astronomical images unfeasible. A critical challenge in this "big data" era is the increasing contamination of images by linear streaks produced by artificial satellites and space debris. These artifacts obscure scientific targets, affect data quality control, and complicate the monitoring of Near-Earth Objects (NEOs). While previous methods have utilized Hough or Radon transforms, or synthetic tracking, there is a need for an automated, scalable solution capable of detecting, characterizing, and cross-identifying these streaks within large datasets to support both planetary defense and space situational awareness.

Methodology
The authors present StreakMind, an end-to-end automated pipeline designed to detect linear streaks in ground-based astronomical images, refine their geometry, associate detections across frames, and cross-identify them with known orbital objects. The system was developed and tested using data from the La Sagra Observatory (MPC code L98), acquired with a Celestron C14+Fastar telescope and an SBIG ST-10 3 CCD camera.

The pipeline architecture consists of the following stages:

Data Preparation and Stratification:
- The dataset comprises 2,055 real FITS images (acquired in 2019) and 280 synthetically generated images.
- Real images were manually inspected, identifying 765 streaks. The distribution was highly skewed toward short streaks.
- To address class imbalance, synthetic streaks were injected into real images using custom Python scripts. These simulations accounted for Point Spread Function (PSF) convolution, random orientations (modeled by a Gaussian distribution), and varying brightness levels.
- A stratified sampling strategy was employed to ensure training, validation, and test sets contained proportional representations of "no-streak," "short-streak," and "long-streak" images.
Detection Model:
- The core detection engine utilizes a YOLO11 (You Only Look Once) model configured for Oriented Bounding Boxes (OBBs).
- The model was pretrained on the DOTAv1.0 dataset and subsequently fine-tuned on the hybrid real-synthetic astronomical dataset.
- Input images were converted from FITS to PNG format, with astrometric coordinates translated into pixel-based OBB annotations.
Post-Processing and Refinement:
- Stellar Filtering: A catalogue-driven filter using Gaia DR3 data removes false positives caused by diffraction spikes from saturated stars.
- Photometric Elongation: Since the initial OBBs often underestimate the true length of faint streaks, a photometric pre-analysis extends the OBBs longitudinally. This involves computing a 1D flux profile along the streak axis and extending the box until the signal drops below a dynamic threshold ( $I_{bg} + k\sigma$ ).
- Endpoint Estimation: The four corners of the refined OBB are clustered to determine stable streak endpoints and the center.
- Inter-Frame Association: Detections across consecutive frames are associated based on geometric extrapolation (velocity estimation) and temporal intervals. The system tolerates up to two missed frames to account for transient observational gaps.
Identification and Database Integration:
- Detected streaks are converted into Minor Planet Center (MPC)-style observation records.
- Cross-Identification: Records are submitted to the Project Pluto Sat_ID service to match against external ephemerides.
- Confidence Scoring: A Gaussian-based confidence model evaluates matches based on angular offsets, penalizing large minimum offsets and weighting candidates relative to the best match.
- Database Export: All results are stored in a normalized SQLite database, linking observations to observatory metadata and satellite identifiers.

Key Results

Performance Metrics: On the held-out test set, the model achieved a precision of 94% and a recall of 97% under a strict Intersection over Union (IoU) threshold of 0.8.
Detection Capabilities: The system reliably detected faint streaks and performed robust geometric reconstructions. The photometric elongation step proved essential for accurately capturing the full extent of long streaks.
Artifact Suppression: The Gaia-based filtering successfully eliminated 77% of false positives caused by stellar diffraction spikes.
Database Integration: The pipeline successfully generated standardized MPC records and cross-identified artificial objects, demonstrating the feasibility of integrating AI inference directly into a structured database for large-scale analysis.

Significance and Claims
The paper positions StreakMind as a significant step toward efficient, automated analysis of orbital-object contamination in astronomical surveys. The authors claim the system offers:

Scalability: It processes entire nights of observations in minutes, a task that would require hours of manual inspection.
Reproducibility: The automated generation of standardized records and confidence scores removes subjective bias and manual cataloguing overhead.
Robustness: The combination of deep learning with geometric and photometric refinement allows for consistent detection across varying streak lengths and signal-to-noise ratios.
Foundation for Future Work: While the current implementation focuses on artificial satellites and debris, the authors note that the pipeline architecture is instrument-agnostic and designed to be extended to NEO identification (via Skybot and JPL Horizons) and adapted to other observatories via transfer learning.

The authors maintain a modest scope, clarifying that the current work is limited to artificial objects and that dedicated photometric analysis (for magnitude estimation) and full NEO identification are planned for future iterations.