11

According to a comment by user @J:

With that said, other than aesthetics, machine vision algorithms are wildly more advanced today than in the past - strategies to remove passing satellites don't really seem to be anything other than trivial. It's annoying because it requires land based observatories to correct for the new noise, but it doesn't seem anything fundamentally more disruptive than that - a mild annoyance.

Is it true that machine learning (and other) algorithms are so good these days that you can automatically eliminate the impact of any satellites passing overhead? Perhaps not the ideal solution for a casual astronomer but perhaps easy to do for a professional observatory?

uhoh
  • 31,151
  • 9
  • 89
  • 293
JonathanReez
  • 547
  • 3
  • 12
  • 1
    I'll add that the context of that comment was strictly about UV-VIS-NIR imaging. The radio pollution created by Starlink is a different matter altogether and I think it's worth stating that I was not trivializing that problem (ie; Starlink RF frying high-gain ground based receivers, etc). That's a more serious problem. – J... Apr 20 '21 at 20:22
  • 5
    @J's comment suggests a profound misunderstanding of the purpose of obtaining astronomical data, which is to obtain data we do not already know, not to make pretty pictures that casual non-scientists think look fine. – Peter Erwin Apr 20 '21 at 20:32
  • 2
    @PeterErwin care to expand this into an answer? – JonathanReez Apr 20 '21 at 20:43
  • Great question! These are different but related: How do observational astronomers manage streaks and other artifacts from objects in Earth Orbit? and Do large telescopes, especially plans for the LSST, avoid saturation artifacts from the brightest stars and planets? If so, how? – uhoh Apr 21 '21 at 00:58
  • @PeterErwin Whether or not the data captured is in the visual part of the electromagnetic spectrum isn't all that relevant per se though, is it? Noise filtering algorithms don't have any wavelength-specific restrictions AFAIK. What does seem relevant though, is the durations of the astronomic events being measured VS satellite traversal time; but yeah, an answer with a specific example of such an issue would be quite welcome I think. – Will Apr 21 '21 at 10:05
  • @Will isn't radio-vs-visible relevant because Starlink satellites are designed to emit radio signals? – user253751 Apr 21 '21 at 10:21
  • @user253751 True, that could be relevant, but I suspect that part of the spectrum is already very noisy without Starlink, plus user J.... already clarified that aspect in the first comment above ↑, so to follow up on that with phrasing suggesting "a profound misunderstanding" seems needlessly dismissive to be honest. – Will Apr 21 '21 at 10:50
  • @PeterErwin Where, may I ask, did I make reference to pretty pictures or say anything at all to give you that impression? Designing, building, and extracting scientific data from imaging systems is one of the many things I do professionally. I'm actually amused at your accusation, to be honest. I'm happy to learn of specific measurements that have no known practical solutions to the noise and light pollution introduced by Starlink, of course. I'm sure in many cases it would be expensive to engineer mitigations, in others probably less so. – J... Apr 21 '21 at 12:58
  • @J it's implicit in the idea that machine learning algorithms can "remove passing satellites". In terms of making the overall image look more "natural" nicer to the human eye, sure. In terms of recovering the actual data hidden by the satellite track, it's an impossibility. – Peter Erwin Apr 21 '21 at 13:17
  • @PeterErwin I suggested nothing about machine learning. OP erroneously inferred that. I was speaking of machine vision, and algorithms or hardware level modifications to allow spatial and temporal binning, etc - removing offending time slices from long exposure integrations, that sort of thing. I also retracted the word "trivial" in my next statement in the discussion that comment was lifted from - because it would represent expensive and extensive modifications to the imaging systems. Conceptually not terribly difficult to solve, but quite expensive. – J... Apr 21 '21 at 13:24
  • @PeterErwin Where learning could be used would be restricted to detection and automated discarding of spoiled data, of course. Obviously there is no recovering lost data, but if you're pointing a camera at a slice of the sky for hours and you lose five minutes of data in a subset of the image area because of a few satellites there's no fundamental reason why you could not simply ignore that data and integrate over the remaining period when the area is clear rather than naively integrating over the whole period and ending up with all of your data being useless. – J... Apr 21 '21 at 13:31
  • @PeterErwin How practical those mitigations would be would depend entirely on the imaging technology in use in each specific instrument. Certain imaging technology classes would be less amenable to certain strategies over others, but I feel that many would generally have some type of engineer-able solution. – J... Apr 21 '21 at 13:35
  • @J Ah, OK, good to know what you were intending, and also good to know that you know the satellite-affected data must be identified and thrown away. What you are talking about doing is pretty second-nature to astronmers, and is fine (except for the lost time) if you are observing something unchanging on the time scale of the observations. – Peter Erwin Apr 21 '21 at 15:21
  • What's missing is consideration of the effects on time-domain data, where an observation of a particular location/object at a particular time is scientifically meaningful. (E.g., eclipses and occultations, potential positions of moving objects like asteroids, and unexpected transients of all different kinds.) – Peter Erwin Apr 21 '21 at 15:41
  • @PeterErwin Indeed, hence my original assertion that this largely more of an annoyance than a total handicap, although certainly there will be certain types of observation that are more severely impacted, as you say. Where the automation would come in would be in the identifying and throwing away bit, surely because at some point grad students will run out of patience having to sift by hand. – J... Apr 21 '21 at 17:11

2 Answers2

4

No, automated algorithms won't ever be able to completely eliminate the effect of passing satellites.

  1. Some of the light reflected from the satellites into the atmosphere will then scatter and cause some amount of light pollution. The pollution isn't limited to just in the direction of observing. See: https://photo.stackexchange.com/questions/41722/can-you-photograph-the-milky-way-with-a-full-moon-out.

  2. Whenever automated algorithms remove the presence of satellites from images, some of the underlying data will be removed as well.

  3. As the number of satellites increases, eventually it will render some data for stellar occultation light curves unusable as the occultations from satellites coincide with the occultations from asteroids. See: Is 486958 Arrokoth (2014 MU69 aka Ultima Thule) the only solar-system object determined to be binary by occultation?. This particular effect will occur even if the satellites have zero reflectivity.

Elon Musk says:

I am confident that we will not cause any impact whatsoever in astronomical discoveries, zero [...] That's my prediction, we will take corrective action if it's above zero.

But he is confident in an impossibility. We can minimize the effect by reducing satellite reflectivity, but we can't make the effect zero, even with advanced algorithms.

Note: The above list is not meant to be comprehensive.

Connor Garcia
  • 16,240
  • 4
  • 43
  • 96
  • we will take corrective action if it's above zero I wonder what Elon had in mind and I'm curious if SpaceX would, or could, ever accommodate "making space" for specific observations. As in, having the swarm make minor orbital adjustments to stay clear of some area of sky within whatever solid angle is under observation. I think this could probably be done up to some limit depending on how much RCS fuel the satellites have for their 4-year life. – J... Apr 21 '21 at 23:20
  • 2
    @J... Re: what Elon has in mind, my money would be on either him no caring, or something in the line of "once I get us on Mars we'll have even better skies for your telescopes" – Luris Apr 22 '21 at 12:37
  • 1
    @Luris which is of course true in the long run. If we manage to build a lunar/Mars colony (which is a major step towards colonizing space, which is in turn the end goal of human civilization), space observations will be a no-brainer. – JonathanReez Apr 22 '21 at 17:30
0

No, you can't perfectly subtract out satellite trails from images. The root problem is that astronomical cameras are essentially photon counters. Any time you are counting, you are limited by Poisson statistics. So, even if you have a fantastic model that says a Starlink satellite will deposit 10,000 photons in a pixel, Poisson statistics say the actual number of photons from the satellite will be 10,000 plus or minus 100. You can subtract off that 10,000 signal, but the 100 noise remains.

This is also why we can't use telescopes during the day. You could make a model of how bright you think the daytime sky should be and subtract it off. But the Poisson noise from all those photons remains.

I.P. Freeley
  • 181
  • 1