SDM: A Two-Stage Coarse-to-Fine Pipeline for Predicting Full-Catalog Satellite Streak Photometry, Astrometry, and Detector-Count Contamination in Zwicky Transient Facility

The night sky, once the domain of stars and distant galaxies, is now crisscrossed by a silent, moving traffic of human-made objects. Since the dawn of the space age, thousands of satellites and pieces of orbital debris have accumulated in low-Earth orbit, circling the planet faster than any bird or plane. For astronomers using powerful telescopes to scan the heavens for fleeting cosmic events, these objects are a growing nuisance. When a satellite passes in front of a telescope's view during a long-exposure photograph, it leaves behind a bright, thin streak that can wash out the faint light of a distant supernova or a near-Earth asteroid. This problem is not just about a few stray lines; it is a contamination issue that threatens the quality of data from the world's most sensitive sky surveys, potentially hiding the very discoveries scientists are trying to make.

To manage this, astronomers need to know not just when a satellite will cross their view, but exactly how bright it will appear and where it will land on their camera sensor. While tools exist to predict the path of an object, they have historically struggled to predict its brightness with the precision needed for modern science. A new study introduces a tool called the Satellite Detection Model, or SDM, which fills this gap. By combining orbital mechanics with a detailed simulation of how sunlight reflects off space junk, the researchers have built a system that can take a list of every known satellite and piece of debris, predict which ones will cross a specific telescope's view, and generate a synthetic image showing exactly how those crossings will look. This allows astronomers to anticipate contamination before they even point their telescopes, turning a chaotic problem into a predictable one.

The researchers behind this work focused on the Zwicky Transient Facility, a survey that uses a large telescope in California to photograph vast stretches of the sky every night. Between late 2019 and the end of 2021, automated searches of the data from this facility identified roughly 73,000 separate events where a satellite glint or streak appeared in the images. These events polluted about 3.6 percent of all the science images taken during that period, involving more than 300 distinct objects. As new mega-constellations of satellites are launched in the coming years, this number is expected to grow, making the ability to predict these intrusions more critical than ever. The challenge is that knowing an object will pass through the field of view is only half the battle; knowing how bright it will be, and whether it will be visible at all, depends on complex factors like the angle of the sun, the distance of the object, and whether it is currently in the Earth's shadow.

To solve this, the team developed a two-stage process that acts like a highly efficient filter. The first stage is a rapid, coarse check that looks at the entire catalog of known space objects—tens of thousands of them—and quickly discards the vast majority that will never come close to the telescope's view. This step is computationally cheap and fast, using a simplified calculation to rule out objects that are too far away or in the wrong part of the sky. By doing this first, the system avoids wasting time on objects that don't matter. The second stage is where the detailed work happens. For the small number of objects that survive the first filter, the system performs a much more precise calculation, updating the position of the satellite and the sun every tenth of a second. It calculates the exact distance between the telescope and the satellite, the angle at which sunlight hits the object, and how much of the object is illuminated versus how much is in shadow.

Using these precise geometric details, the model treats each satellite or piece of debris as a sphere that reflects light in a predictable way. It calculates how much light reaches the telescope and converts that into a brightness value, simulating exactly how many pixels on the camera sensor would be lit up. The system then paints these results directly onto a digital canvas that matches the size and shape of a real telescope image. The result is a synthetic photograph that shows every satellite streak that will appear during a specific exposure, complete with its correct position, length, and brightness. This allows researchers to see a "what-if" scenario before the telescope even takes the picture, helping them understand how much of their data might be compromised.

The study tested this model by running it against a simulated exposure window, generating a synthetic image that showed dozens of predicted streaks. The output included a detailed list of every object that crossed the view, noting its type—whether it was a rocket body, a piece of debris, or an active satellite—and its estimated brightness. The model proved capable of distinguishing between objects that would be bright and those that would be invisible because they were in the Earth's shadow. For instance, the simulation correctly showed that some objects, even if they passed very close to the telescope, would appear completely dark because they were not illuminated by the sun at that moment. This ability to account for shadows is a significant improvement over simpler models that might assume an object is always visible if it is close enough.

However, the researchers are careful to note that their current model is a simulation based on assumptions rather than a final, perfect prediction. Because the physical size and reflectivity of most space objects are not known precisely, the system assigns them likely values based on their names and types. A piece of debris might be modeled as a small sphere with a certain reflectivity, while a large satellite might be treated as a bigger, more reflective sphere. This means that while the model is excellent at showing the general pattern of contamination and the relative brightness of different objects, the exact brightness of any single satellite in a real image might differ from the prediction. Furthermore, the model currently treats all objects as smooth, round spheres, which does not account for the complex shapes of real satellites or the way they might tumble and flash as they spin. These "glints," which can cause sudden spikes in brightness, are not yet fully captured by the system.

Despite these limitations, the tool represents a major step forward in managing the clutter of low-Earth orbit. By providing a way to generate a full-catalog, per-exposure prediction of satellite contamination, the Satellite Detection Model gives astronomers a powerful new way to plan their observations. They can use it to forecast which parts of the sky will be most polluted on a given night, or to automatically flag the pixels in an image that are likely to be affected by a satellite streak so that they can be cleaned up later. The researchers plan to take the next step by comparing their simulated images against real data from the Zwicky Transient Facility to see how closely the predictions match reality. If the model holds up, it could become a standard tool for scheduling observations and protecting the integrity of future sky surveys, ensuring that the view of the universe remains as clear as possible even as our own orbit becomes increasingly crowded.