A telescope that tracks a satellite can leave stars trailing across its image; one that follows the stars can turn the moving satellite into a streak. That complicates a basic measurement: how bright is the target? Aperture-X addresses the problem by combining several physically motivated ways of measuring light in a machine-learning model.[1][2][3]
Published in Astronomy and Computing, the method uses XGBoost to combine four measurements into a visual-magnitude estimate. The attraction is that each input has an identifiable role in the measurement, giving researchers a way to examine which features the model uses when an image is difficult.[1][2][3]
Four views of the same light source
A simple aperture measurement adds the light inside a chosen region and estimates the background around it. The region must be wide enough to capture the target without including too much unrelated light. A fixed circle can work well for a compact source but poorly for a smeared or elongated one.
Aperture-X combines a fixed-radius measurement, a full-width-at-half-maximum approach, a depth-first-search method with an interquartile-range filter, and an adaptive Gaussian point-spread-function estimate. Each captures a different aspect of the light distribution. The learned model uses their combined information rather than selecting one measurement rule for every image.[2][3]
Training used about 1.2 million synthetic patches, including 868,000 star and 293,000 satellite examples. Under simulated stray light, the authors report reductions in mean absolute error of up to 41% for stars and 48% for satellites relative to the study’s analytical and learning-based comparisons. These results concern the synthetic tests, with real-image transfer assessed separately.[2][3]
The evaluation uses distinct kinds of data
| Experiment | Data | Reported purpose or result |
|---|---|---|
| Stray-light stress test | Synthetic stars and satellites | Up to 41% and 48% lower mean absolute error |
| Transfer to real-star photometry | About43,000 stars | Tests adaptation from simulated to real observations |
| Satellite-like proxy test | 44,000 injected point sources | Tests known sources within real image backgrounds |
| Satellite outlier measure | Simulated satellite examples | Fewer than 0.1% exceed 0.3 magnitude error |
Brightness error has its own units
Astronomical magnitude is logarithmic: a difference in magnitude represents a ratio of light, rather than a linear change. The authors define an outlier as an estimate whose error exceeds 0.3 magnitude and a characterisable estimate as one with error below 0.1 magnitude. Those thresholds describe photometric error, not a target’s position or distance.[2][3]
For satellite observation, a brightness history can help researchers investigate changes in orientation or reflected light alongside positional tracking. Better photometry gives those analyses a more reliable input. It is especially useful to know when a method remains stable in glare or with unusual target shapes, because those are conditions in which a neat average can hide difficult individual observations.
The simulation-to-reality step is therefore central to the work. Synthetic images provide known answers for training and controlled tests; real backgrounds introduce conditions the model must handle beyond that setup. By separating the real-star, simulated-satellite and injected-source experiments, the study makes it possible to see which part of that transfer has actually been tested.
Sources & further reading
- Aperture-X: Physics-informed aperture feature learning for robust photometry Astronomy and Computing · Accessed 25 Oct 2026
- Aperture-X: Physics-Informed Aperture Feature Learning for Robust Photometry SSRN preprint · 26 Feb 2026 · Accessed 25 Oct 2026
- Aperture-X: Physics-Informed Aperture Feature Learning for Robust Photometry — author abstract ResearchGate / author preprint record · Accessed 25 Oct 2026