🔭 astrophysics

Improved Identification of Satellite Trails in ACS/WFC Imaging Using a Modified Radon Transform

This paper introduces a modified Radon Transform algorithm integrated into the \texttt{acstools} package to effectively identify faint satellite trails in ACS/WFC imaging, revealing that while contamination rates have doubled over the past two decades, trail brightness has remained stable.

Original authors: David V. Stark, Norman Grogin, Jenna Ryon, Ray Lucas

Published 2026-02-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: David V. Stark, Norman Grogin, Jenna Ryon, Ray Lucas

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 Problem: Space Junk in the Sky

Imagine you are trying to take a beautiful, long-exposure photograph of a starry night sky. You want to capture the faint glow of distant galaxies. But, suddenly, a bright streak cuts across your photo. It's not a shooting star; it's a satellite (like the International Space Station or a Starlink satellite) passing overhead.

For astronomers using the Hubble Space Telescope (HST), this is a growing nightmare. As more and more satellites are launched into orbit, Hubble's photos are getting "ruined" more often. These streaks look like scratches on a painting, and they make it hard to study the real science behind the image.

The Old Way: Finding a Needle in a Haystack

Previously, astronomers tried to find these streaks using two main methods:

Looking by eye: Humans would stare at thousands of images and draw lines around the streaks. This is slow and boring.
Old computer algorithms: These were like trying to find a needle in a haystack by looking for anything "sharp." The problem? The "haystack" is full of other sharp things: bright stars, galaxies, and the diffraction spikes (those cross-shaped patterns that bright stars make in telescope photos). The old computers got confused, thinking a bright star was a satellite, or missing a faint satellite because it was too dim.

The New Solution: The "Median Radon Transform" (MRT)

The authors of this paper invented a new, smarter way to find these streaks. They call it the Median Radon Transform (MRT).

Here is how it works, using a simple analogy:

The "Sum" vs. The "Median" Analogy
Imagine you are walking through a crowded room (the image) and you want to find a line of people holding hands (the satellite trail).

The Old Method (Sum): You add up the "loudness" of everyone you meet. If there is one person screaming (a bright star) in the middle of the room, your total "loudness" goes through the roof. You might think the whole room is a line of people just because of that one screamer.
The New Method (Median): Instead of adding up everyone's volume, you ask, "What is the middle volume of the people in this line?" If there is one person screaming, but everyone else is whispering, the "middle" volume is still a whisper. The screamer gets ignored.

By using the Median instead of the Sum, the new algorithm can ignore the bright stars and galaxies that usually confuse the computer. It only cares about the continuous "line" of pixels that makes up the satellite trail. This allows it to spot trails that are much fainter than the background noise—like spotting a whisper in a library.

How They Tested It

The team tested their new "satellite hunter" on real Hubble data:

Success Rate: They compared their computer's findings to a list of trails found by humans. The computer found 85% of the trails the humans found.
False Alarms: It made mistakes only 2.5% of the time (after ignoring the obvious cross-shaped star patterns).
The "Corner" Problem: The algorithm struggles a bit in the corners of the image. Imagine trying to draw a straight line across a square piece of paper; if you start in the corner, the line is short. Short lines are harder to distinguish from random noise or a group of stars that just happen to look like a line.

The Big Discovery: Space is Getting Busier

Using this new tool, the team looked at 20 years of Hubble data (from 2002 to 2022). They found two major things:

More Traffic: The number of satellite streaks hitting Hubble's photos has doubled in the last 20 years. As we launch more satellites, the sky is getting more crowded.
No Brighter Traffic: Interestingly, the satellites aren't getting brighter or dimmer over time. They are just showing up more often.

Why This Matters

This new tool is now available for free to all astronomers (it's part of a software package called acstools). It acts like a high-tech "eraser" that can automatically find and hide these satellite streaks, cleaning up the data so scientists can focus on the universe, not the traffic.

In a nutshell: The astronomers built a smarter filter that ignores the "loud" stars to find the "quiet" satellite trails. It works great, saves time, and confirms that our sky is getting increasingly crowded with human-made objects.

1. Problem Statement

Astronomical imaging data from the Hubble Space Telescope (HST), specifically the Advanced Camera for Surveys (ACS) Wide Field Channel (WFC), is frequently contaminated by linear artifacts. The most prevalent of these are satellite trails, caused by the increasing population of human-made satellites in Low Earth Orbit (LEO). Other linear artifacts include cosmic rays, scattered light, charge bleeding, and diffraction spikes.

Existing automated detection methods face significant challenges:

Sensitivity: Many algorithms fail to detect trails with mean brightness significantly below the background noise level.
Robustness: Standard detection methods (e.g., edge detection or standard Radon Transforms) are easily confused by bright astronomical sources (stars, galaxies), which can produce false positives or mask real trails.
Evolution: With the rapid increase in satellite density (e.g., Starlink), the contamination rate is rising, necessitating more efficient and robust automated masking tools.

2. Methodology

The authors introduce a new algorithm, findsat_mrt, which utilizes a Modified Radon Transform (MRT) to detect linear features.

The Median Radon Transform (MRT)

The standard Radon Transform (RT) calculates the sum of pixel values along all possible linear paths ( $\theta, \rho$ ) through an image. While effective for isolated linear features, the RT is highly susceptible to bright point sources (stars), which elevate the transform values along paths passing through them, obscuring faint trails.

The authors replace the summation with the median:
$MRT(\theta, \rho) = \tilde{f}(x, y)_L$
Where $\tilde{f}$ is the median of the function $f$ over a path $L$ .

Advantage: The median is resistant to outliers. In a background-dominated image, a single bright star will not significantly alter the median value along a path, whereas a continuous linear trail (even if faint) will shift the median. This allows the MRT to detect trails with brightness well below the noise floor while ignoring bright stars.

Algorithm Workflow

Preprocessing:
- Input: ACS/WFC flc images (calibrated).
- Bad pixels are masked (set to NaN).
- Crucial Step: Cosmic ray flags are not removed, as they can partially flag satellite trails, hindering detection.
- Rebinning: Images are 2x2 rebinned to reduce computation time by an order of magnitude without losing sensitivity (trails are typically several pixels wide).
- Median background subtraction.
MRT Calculation:
- Computed using skimage.transform.radon with the summation replaced by a median.
- Performed on 24 cores; a single 2048x1024 image takes ~10 seconds.
- An uncertainty map is generated based on the standard error of the median.
Source Identification & Filtering:
- Detection: A source finder (photutils) identifies peaks in the MRT space using kernels of varying widths (3, 7, 15 pixels) to account for different trail widths.
- Refinement: Candidates are rotated to a horizontal orientation. A 1D cross-section is created, background is modeled via polynomial fit, and a Gaussian is fitted to determine width and peak intensity.
- Persistence Check: To distinguish real trails from chance alignments of stars, the trail is divided into $N$ sections. A "persistence score" is calculated based on how many sections maintain a Signal-to-Noise (S/N) > 3. Trails with a score < 0.5 are rejected.
- Multi-chip Validation: For ACS/WFC (two chips), the algorithm attempts to match trails crossing the chip gap. If a trail is detected on both chips with consistent angles, it is confirmed as robust.

3. Key Contributions

Algorithm Development: Creation of findsat_mrt, a Python-based tool integrated into the acstools package (v3.6.0+).
Methodological Innovation: The application of the Median Radon Transform specifically to astronomical imaging, demonstrating its superiority over the standard Summed Radon Transform in crowded fields.
Comprehensive Validation: Extensive testing against visual inspections (Frontier Fields) and large-scale statistical analysis of 20 years of HST data to define reliability boundaries.

4. Results

Performance Metrics

Sensitivity: The MRT can detect trails with a median brightness as low as 0.1 times the background noise (for long trails crossing the chip center). This is roughly an order of magnitude more sensitive than the previous detsat routine.
Recovery Rate: When compared to visual inspections of Frontier Fields data, the algorithm recovers 85% of identified trails (strict definition) or 88% (relaxed definition).
False Detection Rate:
- 2.5% after removing diffraction spikes (which are easily filtered by angle).
- ~10% in general parameter space if diffraction spikes are not filtered and data includes crowded fields or image corners.
Failure Modes:
- Image Corners: Sensitivity drops significantly for trails < 500 pixels long (corners of the detector).
- Crowded Fields: In fields with high source density (defined by a skewness parameter $b > 50$ ), chance alignments of stars can mimic trails.
- Diffraction Spikes: The algorithm detects these but cannot inherently distinguish them from trails without angle-based filtering.

Evolution of Satellite Contamination (2002–2022)

The authors analyzed ACS/WFC data over two decades to track contamination trends:

Frequency: The rate of satellite trail contamination has increased by a factor of ~2 between 2002 and 2022.
Brightness: There is no clear systematic evolution in the typical brightness of satellite trails over time.
Data Quality: The study identified that restricting analysis to fields with $b < 50$ (low source density) and trail lengths $L > 500$ pixels yields a clean sample with a false positive rate of ~10%.

5. Significance and Recommendations

Critical for Future Data: As the number of LEO satellites grows, automated masking is essential for preserving the scientific utility of HST data.
User Guidelines: The authors recommend:
- Ignoring trails at angles corresponding to diffraction spikes ( $\sim0^\circ, 90^\circ, 180^\circ$ ).
- Manually inspecting trails in crowded fields ( $b > 50$ ) or short trails ( $L < 500$ ).
- Using multi-epoch data to distinguish persistent artifacts (like glints) from stochastic satellite trails.
Future Improvements: The authors plan to apply the algorithm to distortion-corrected (drc) images to increase trail length and sensitivity, and to explore faster calculation methods (Fast Radon Transform) to reduce computational cost.

In conclusion, the findsat_mrt algorithm represents a significant advancement in automated artifact removal for HST data, offering high sensitivity to faint trails and robustness against bright sources, thereby enabling cleaner scientific analysis of the ACS/WFC archive.