What is synthetic tracking, and why would a 35 cm Earth imager be 10-30x better than Pan-STARRS or LSST for interstellar asteroid discovery?

The ArXiv preprint Technical Note: Asteroid Detection Demonstration from SkySat-3 B612 Data using Synthetic Tracking is an interesting read! The abstract says:

We report results from analyzing the B612 asteroid observation data taken by the sCMOS cameras on board of Planet SkySat-3 using the synthetic tracking technique. The analysis demonstrates the expected sensitivity improvement in the signal-to-noise ratio of the asteroids from properly stacking up the the short exposure images in post-processing.

This is a system optimized for short (20 ms) exposures of Earth's bright, sunlit surface using CMOS imagers (not CCDs), being pointed into space during the eclipse phase of its polar orbit, and imaging dim, deep space objects. For this test asteroids with a visible magnitude of roughly +14 to +16 were tested, but according to section 6.2, the same 35 cm telescope with a small redesign of the hardware would have a limiting magnitude of about +21!

The following caught my eye:

6.3 Future Directions in Asteroid Searches

Recently Pan-STARRS announced the first discovery of an interstellar asteroid. This came after ~10 years of Pan-STARRS operation. Early estimates are that the discovery rate would only increase slightly in the near future, even with telescopes like LSST. The principal reason is that interstellar asteroids move very rapidly compared to normal NEOs, thus producing a streaked image with PanSTARRS and LSST. Synthetic tracking would be able to increase the discovery rate by 10~30, even when using relatively small telescopes such as the 35cm aperture SkySat-3 telescope.

Question: What is synthetic tracking, and why would a 35 cm Earth imager be 10-30x better than Pan-STARRS or LSST for spotting interstellar asteroids? Is there something fundamental about the Pan-STARRS system or the Large Synoptic Survey Telescope that is unalterable for use in this specific application?

The usual approach to tracking a dim, moving object is to move the camera to follow a predicted track while making a long exposure.

Synthetic tracking starts with a series of sequential quick exposures. If the camera were mechanically tracking, you could just overlay and add them. (The best statistical method for combining is a little more complicated, but you get the idea)

If the camera was not tracking, you can still do this by computationally moving the images before combining. You compute how the image would be displaced if the tracking was perfect and apply that to synthesize each proper image, then stack. Hence “synthetic tracking”.

One big advantage for asteroid searches is that you can synthetically track a lot of possible motions with a single set of images. Mechanical tracking can only track one target motion at a time.

To work best, you want a lot of quick exposures. Ideally, you want the exposure time so short that the tracked target(s) move less than one pixel during the exposure. This limits blurring in the synthetic image.

CCDs versus CMOS imagers

Unfortunately, most telescope cameras work in the other limit: they stare for a long time. This is great for mechanically tracking images, but results in streaky synthetically-tracked images. Although the sensor noise sources (photon shot noise and thermal noise) are similar, the CCDs usually used for astronomical cameras also have a readout noise source which grows with frequency. This is fine for long exposures, but makes direct-readout CMOS sensors much more suitable for rapid-sequence readout.

We’re used to observing moving objects as streaks in images, so what’s the problem? Those are bright streaks. Dim streaks are different. They spread the target light over a bunch of pixels, each with their own noise. That reduces the overall SNR of the observation, and makes the streak less visible. Synthetic tracking of sufficiently short images can get right down to the single-pixel noise limit of the sensor, but only if the exposures are short enough.

Numerical example: a 1000 pixel streak in a normal camera with a 10-second exposure has 10 seconds of noise in each pixel. With synthetic imaging, each pixel is e.g. sampled at a particular 10ms slice, so (all other things equal) has $$\sqrt{100}$$ less noise. That allows detection of 10-times-dimmer asteroids.

N.B. “Interstellar asteroid” here doesn’t mean “out really far in interstellar space”. It means “moving like a bat out of hell because it’s fallen in from interstellar space instead of being in solar orbit”. They’re dim, but they’re mostly hard to see with fixed imagers because they move so fast.

• Thanks! added a bit on ccd vs direct-readout sensors and the noise calculation – Bob Jacobsen Jul 7 '19 at 0:27
• They currently have CCD cameras. – Bob Jacobsen Jul 7 '19 at 0:43
• It is also helpful to note the huge scale of the detector arrays, and the absolute impracticality of swapping the CCD version for a new CMOS version for some observing sessions, e.g. for LSST this and this This addresses the "something fundamental and unalterable" aspect of the question. – uhoh Jul 7 '19 at 1:00
• fyi I've just asked Why exactly do “CMOS astrophotographers” prefer CMOS sensors? – uhoh Jul 7 '19 at 2:42