🔭 astrophysics

Streak detection in the VST/OmegaCAM archive using deep learning

This paper presents a robust deep-learning pipeline that successfully detects and classifies satellite and space debris streaks in the VST/OmegaCAM archive, achieving high accuracy on real data and uncovering a significant population of uncatalogued objects to enable large-scale space debris monitoring without dedicated observing time.

Original authors: Elisabeth Rachith, Stephan Hellmich, Vincent Fiszbin, Belén Yu Irureta-Goyena, Andrew Price, Jean-Paul Kneib

Published 2026-06-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Elisabeth Rachith, Stephan Hellmich, Vincent Fiszbin, Belén Yu Irureta-Goyena, Andrew Price, Jean-Paul Kneib

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-definition movie screen used by astronomers to study stars and galaxies. Unfortunately, this screen is getting cluttered. Every night, invisible "ghosts" in the form of satellites and space junk fly across the view, leaving behind bright, scratch-like lines (streaks) that ruin the picture.

This paper describes a new, automated system built by researchers at EPFL in Switzerland to find, sort, and study these scratches in a massive library of old telescope photos.

Here is how they did it, explained in everyday terms:

1. The Problem: A Library of Messy Photos

The researchers used the VST/OmegaCAM, a giant camera on a telescope in Chile. It has been taking photos of the sky for over a decade, creating a library of hundreds of thousands of images.

The Issue: These photos are full of "streaks" from space debris. While astronomers usually want to remove these to see the stars, the researchers realized these streaks are actually valuable. They are free data about space junk that we didn't have to pay to collect.
The Challenge: There are too many photos to look at by hand. Also, not every line is a piece of space junk; some are just camera glitches, bright stars bleeding light, or weird patterns caused by the telescope's filters.

2. The Solution: A Two-Step Robot Team

The team built a "robot team" using Artificial Intelligence (AI) to handle the work. Think of it as a two-person inspection crew:

Step 1: The "Sniffer" Dog (The Detector)

What it does: This AI looks at small chunks of the photos and screams, "I see a line!"
How it works: It uses a special trick called a "Hough Transform" (which is like a mathematical way to spot straight lines) combined with a deep learning brain. It was trained on a mix of real photos and fake, computer-generated streaks that looked exactly like real space junk.
The Result: It is very good at finding lines. It catches over 95% of the real streaks, even the faint ones. However, like an over-enthusiastic dog, it also barks at things that aren't streaks (like star glare or camera noise).

Step 2: The "Judge" (The Classifier)

What it does: This second AI acts as a strict judge. It takes the lines the "Sniffer" found and asks, "Is this actually space junk, or is it just a glitch?"
How it works: It looks at the shape and texture of the line. It learned to recognize specific "fake" patterns, like the cross-shaped spikes from bright stars or the rainbow fringes from camera filters.
The Result: It is incredibly accurate. It throws out almost all the false alarms (96% of them) while keeping almost all the real space junk (97%).

3. The Big Discovery: The "Unknowns"

After training their robots, the team ran them through one full year of the telescope's photo archive (over 1.2 million individual image panels).

The Findings: They found 25,335 streaks.
The Surprise: About 20% of these streaks did not match any known satellite or piece of debris in the official government catalogues.
What this means: The telescope found "ghosts" that nobody knew existed. These are likely small, untracked pieces of space junk that are too faint for radar to see but bright enough for this giant camera to catch.

4. Why This Matters

Free Data: We don't need to build new satellites to track space junk. We can just use the photos astronomers are already taking for free.
Safety: By finding these "unknown" objects, we get a better picture of how crowded space really is, which helps protect future satellites and astronauts.
Efficiency: The system is fast and automated. It can process thousands of images in a time it would take a human years to do.

The Bottom Line

The researchers built a smart, two-step AI system that acts like a high-speed filter for a massive photo library. It successfully separates the "real" space junk from the "camera glitches" and discovered that a significant portion of space debris is currently invisible to our official tracking lists. They proved that we can turn "ruined" astronomical photos into a valuable treasure trove for space safety.

Technical Summary: Streak Detection in the VST/OmegaCAM Archive Using Deep Learning

Problem Statement
Ground-based astronomical surveys, such as those conducted by the VLT Survey Telescope (VST) equipped with the OmegaCAM imager, inadvertently capture streaks from satellites and space debris crossing their fields of view. While these incidental observations offer a unique opportunity to characterize resident space objects (RSOs) without dedicated observing time, the sheer volume of archival data (over 400,000 exposures) and the presence of various image artifacts make manual analysis impossible. Existing detection methods, such as matched filtering or Hough/Radon transforms, often struggle with computational costs, false positives from linear artifacts (e.g., diffraction spikes), or lack the robustness required for large-scale surveys. Furthermore, commercial Space Situational Awareness (SSA) services rarely catalogue debris observations for scientific use, and dedicated surveys are limited by operational costs and narrow tracking capacities. The primary challenge addressed is the development of an automated, scalable pipeline to detect, classify, and characterize these streaks within the OmegaCAM archive to monitor space debris populations and assess data contamination.

Methodology
The authors developed a two-stage deep-learning pipeline designed to process raw OmegaCAM images:

Streak Detection (HT-LCNN):
- The initial detection utilizes an adapted Hough Transform Lookup-based Convolutional Neural Network (HT-LCNN). This architecture integrates a Hough transform within a residual branch of a CNN to enhance the recognition of linear geometric features.
- The model processes 512 × 512 pixel subcrops from raw 16-bit monochrome images to balance contextual information with computational efficiency.
- Post-detection, a "stitching" procedure reconstructs streaks across the full CCD tile by grouping detections with similar Hough parameters ( $\rho$ and $\theta$ ) and selecting the most distant endpoints. This step filters out singleton false positives that do not align with neighboring detections.
- A refinement step aligns the detected streaks horizontally, splits them into segments, and fits Moffat profiles to the cross-sectional intensity to precisely determine the streak center and endpoints.
False Positive Classification:
- To address the high rate of false positives inherent in the detection stage (caused by diffraction spikes, saturation bleeding, filter fringes, and stray light), a secondary classifier based on the VGG6 architecture is employed.
- This classifier takes 144 × 144 pixel cropped images of the refined streaks and outputs a probability score for true positive detection.
- The training dataset was augmented with 224,291 physically simulated streaks generated using Langevin dynamics to model atmospheric scintillation and Moffat profiles for seeing effects, alongside 1,861 manually annotated real streaks.
Astrometric Calibration and Correlation:
- Detected streaks undergo astrometric refinement using Astrometry.net.
- Results are cross-matched with the Space-Track catalogue by propagating Two-Line Elements (TLEs) to the observation time and comparing predicted positions with observed streaks based on angular separation and position angle.

Key Contributions

Pipeline Development: The creation of a robust, automated pipeline combining HT-LCNN for detection and VGG6 for classification, specifically tailored for the 16-bit raw data format of astronomical archives.
Dataset Augmentation: The generation of a large-scale, physically simulated dataset of space debris streaks, incorporating realistic atmospheric effects (scintillation and seeing), to overcome the scarcity of manually annotated real-world data.
Archival Analysis: The application of this pipeline to one year of VST/OmegaCAM observations (1,246,048 CCD frames), representing a significant scale-up in automated space debris monitoring from survey archives.

Results

Detection Performance: On the augmented validation and test sets, the detector achieved F1-scores of 0.966 and 0.958, respectively, detecting over 95% of artificial streaks with a signal-to-noise ratio (S/N) > 4.
Real-World Performance: When applied to real 2023 data, the detection-only precision dropped to 0.783 due to image variability. However, the addition of the VGG6 classifier boosted precision to 0.990 while retaining 97% of true positives and rejecting >96% of false positives.
Catalogue Correlation: Applied to 2023 data, the pipeline identified 25,335 streaks. Approximately 20% of these confirmed detections could not be correlated with any entry in the Space-Track catalogue, suggesting the detection of uncatalogued or unidentified objects.
Contamination Statistics: Over the one-year period, 16.9% of the images contained at least one streak, though only 1.9% of individual CCDs were affected.
Bias Analysis: The pipeline showed a bias against detecting very small Low Earth Orbit (LEO) debris compared to the Space-Track catalogue, attributed to the optical telescope's lower sensitivity to small, faint objects compared to radar systems. Conversely, it successfully detected objects ranging from 359 km (Starlink) to 88,060 km (SL-6 R/B) and objects with ages spanning from 12 days to nearly 65 years.

Significance and Claims
The paper claims that the proposed pipeline demonstrates robust performance on real archival data, successfully uncovering faint, uncatalogued objects that are typically missed by standard tracking methods. The authors highlight the potential of leveraging open-access survey archives for debris monitoring, offering a cost-effective method to derive insights into space object populations (such as rotation rates and lightcurves) without dedicated observing time. The study underscores that no region of the sky is free from satellite-induced optical interference, necessitating precise quantification of streak impacts on astronomical data. While the authors acknowledge limitations in generalizing to all real-world conditions due to the vast diversity of the archive and the difficulty of fully simulating real streak morphologies, they assert that their approach provides a significant step forward in characterizing the space debris population and assessing data contamination.

1. The Problem: A Library of Messy Photos

2. The Solution: A Two-Step Robot Team

3. The Big Discovery: The "Unknowns"

4. Why This Matters

The Bottom Line

More like this