Satellites
Crossing <1 s
LINES DRAWN BY THINGS WE SENT UP.
Not every streak in a difference image is a rock. A satellite in low Earth orbit crosses the entire field of view in under a second, laying a bright line edge to edge through whatever the telescope was actually trying to observe. It is a real object, moving fast, in a sharply defined trail — which is to say it looks exactly like the near-Earth asteroid the pipeline is hunting, right up until you measure how long the line is. The instrument below is invented. The problem is extremely real.
NOCTIS ARRAY does not exist. The satellite problem does — and this page explains it the way a real time-domain difference-imaging pipeline actually has to think about it.
Sunlight reflecting off hardware, a few hundred kilometres up.
A satellite is not glowing. It is a piece of metal and solar panel catching sunlight while the ground beneath it is already in darkness — which is precisely why the worst contamination happens in the hours right after dusk and right before dawn, the same twilight windows a survey most wants for low-airmass observing.
Nothing astronomical moves this fast. That is both the problem and the solution.
Low Earth orbitFast because it is close.
An object at a few hundred kilometres altitude orbits at roughly 7.5 kilometres per second and completes a lap in about ninety minutes. From the ground that translates to angular speeds around a degree per second — five or six orders of magnitude faster than even an extreme near-Earth asteroid. A telescope with a one-degree field sees the whole crossing in about a second.
Not just satellitesRocket bodies and debris too.
The same signature comes from spent upper stages, discarded hardware and collision fragments, of which there are far more than there are working spacecraft. Tumbling objects glint — their reflected brightness flashes as different faces rotate into the sunlight — which turns a smooth trail into a dashed or beaded one and gives an already awkward shape another way to confuse a detector.
The population is growingFast, and by design.
Large communications constellations have added tens of thousands of objects to low Earth orbit within a few years, and the planned totals are larger still. For a wide-field survey this is not an occasional nuisance but a structural change in the data: a meaningful fraction of exposures now carry at least one trail, and no amount of careful scheduling avoids them entirely.
Both are streaks. One of them ends inside the frame.
A near-Earth asteroid at close approach draws a trail of tens of pixels — long enough to be obvious, short enough to have two visible ends. A satellite draws a line across the entire detector. Length is the discriminator, and it is a clean one.
A fast asteroid leaves a mark. A satellite leaves a scar.
Satellites are not a science product here. They are a gate the pipeline has to pass.
The streak branch runs a pretrained three-gate cascade over every elongated detection. The first gate asks whether the trail is real at all; the second, deliberately stricter one, throws out cosmic rays; the third and last asks short or long. Short means an asteroid candidate and continues down the pipeline. Long means satellite, and stops there.
Across 666,811 detections in 503 processed exposures, that final gate has routed 16 detections to LONG_SATELLITE, against 1,057 short asteroid candidates, 2,343 cosmic rays and 419,160 rejections. The satellite count is low for a mundane reason rather than an impressive one: our stamps are cut around catalogued detections in archival fields chosen for reference depth, and a full-frame satellite trail tends to fragment into many separate detections rather than presenting as one clean object. We do not claim this as a satellite census.
The gate earns its place for an honesty reason, not a scientific one. A satellite streak is fast enough to link into a track exactly like an asteroid — but satellites are not in the minor-planet catalogue our novelty filter queries. So a satellite track would come back "no known fast mover at this position" and read as novel: a fabricated discovery, produced by a correctly-functioning pipeline asking the wrong catalogue. The linker is therefore restricted to short-routed candidates only. The short-versus-long gate already told us which was which; the design decision is simply to honour it.
The cascade itself is validated on real data. Run against the two dozen genuine ZTF streak stamps published with the pretrained model, its real-versus-bogus gate accepted 24 of 24, and routed 21 of them as short asteroid candidates — with the remaining three correctly ditched by the cosmic-ray gate as messy or edge-clipped.
This is now a data-quality problem for every wide-field survey on Earth.
Filtering trails out is straightforward. Filtering them out without filtering out the fast-moving asteroids you were looking for, in a sky that carries more of them every year, is not.
A trail damages more than itselfContamination is not local.
A bright line across a frame does not merely occupy its own pixels. It disturbs the background estimate around it, can saturate and bleed along detector columns, and leaves subtraction residuals on either side once the frame is differenced against a reference that has no trail in it. One satellite can degrade a region far wider than the line it drew.
The cut is not freeEvery filter costs real detections.
Any rule aggressive enough to remove satellite trails reliably will also remove some genuine fast movers, because at the extreme end the two populations overlap in exactly the property being cut on. This is the same asymmetry that runs through the whole pipeline: a rejected artefact costs nothing, a rejected real object is gone. Which is why the decision is made by a trained gate that can be measured, rather than by a hand-tuned length threshold nobody can audit.
Worth saying plainly. Satellites are not a flaw in the sky. They are useful infrastructure that happens to sit between telescopes and everything beyond them, and the mitigation work — darkening coatings, orientation control, published ephemerides so observatories can schedule around passes — is real and ongoing. From this pipeline's narrow point of view they are simply a class it must recognise and set aside before it is allowed to say the word "candidate" about anything else.
Go see what the cascade actually routed.
The array, the altitude and the field notes above are invented. The browsers below are not — they open the same working ZTF pipeline whose short-versus-long gate is described in the panel above.
Back to the overview