TBD2_JPL_proposers_day_briefing.pdf

TBD2 Cislunar SSA Track Before Detect Algorithm Evaluation

M. Shao (Michael.shao@jpl.nasa.gov) Russel Trahan (russel.trahan@jpl.nasa.gov)

TBD algorithms are designed to detect moving objects that are not normally detectable in a single image.

© 2025 California Institute of Technology. Government sponsorship acknowledged.

Distribution Statement A: Approved for public release. Distribution unlimited.

mailto:Michael.shao@jpl.nasa.gov mailto:russel.trahan@jpl.nasa.gov mailto:russel.trahan@jpl.nasa.gov mailto:russel.trahan@jpl.nasa.gov

JPL, NASA’s Center for Solar System Exploration

• JPL is known for its spacecraft that flew by, orbited, and landed on other planets in our solar system.

• Our group started in this area looking for Near Earth Asteroids with synthetic Tracking (2014)

• Synthetic Tracking is now deployed in USSF’s GEODSS network of ground-based space surveillance telescopes.

A recently (10/30/2025) discovered Near Earth Asteroid, 2025 UC11. Discovered ~200,000 km from Earth flew by the Earth ~225km above the surface. Closest flyby recorded.

Mars Rover

Jupiter from Voyager 1

Distribution Statement A: Approved for public release. Distribution unlimited.

• 20.2 mag at time of discovery

• Followed by other observers, with orbit calculate its H mag (abs mag), H 27.9mag

• Size (assuming 14% albedo) ~9.2m

• ~3.5 million km from Earth

• Same brightness as a ~1m object at Lunar dist.

Some examples

2021 RF6 Observed at 20210908UT06:43:32 Rate = [4.4,12.7] deg/day ~0.57 as/sec G-mag = 20.1, SNR ~9.8

Tracking Sidereal, streak is too faint to detect

Sky background 20.5mag, sky noise in the image is 22.5 mag

Sensor was a 28cm diameter telescope (11 inch), f/2.2 with a 60 Mpix CMOS, 7.3sqdeg FOV

Sidereal imageSynthetic image

Near Earth Asteroid example:

ZeroV and Synthetic Images NASA Capstone 6U Cubesat lunar NRHO

Feb 4, 2023 Synthetic image 19.5mag Coadding 600 frames With a velocity vec (0.25, -0.07) arcsec/sec

Feb 5, 2023 Synthetic image

18.6 mag Coadding 600 frames With a velocity vec (0.27, -0.04) arcsec/sec

This is a zeroV or normal image 3 15mag stars are visible. (sky ~14mag/as2) Dotted line is the 162 arcsec of motion of the cubesat during the 600 sec obs.

CisLunar SSA/SDA from Solar L1

• Ground based observatories are the workhorse of current SSA activities. But the sensitivity of ground-based telescopes is very poor in daytime.

• Space based observatories don’t have sunlight scattered by the atmosphere, but they still have the phase effect.

• If one wants persistent coverage of all of cislunar space, some (or all) of the SSA satellites have to be outside of the moon’s orbit.

• When the SSA satellite is outside the moon’s orbit, the solid angle one has to search is no longer 4π steradians, it’s much less.

• This then leads to the idea of placing the SSA satellite at ESL1. The downside is a 1m RSOs at a distance of 1.5 million km are much fainter ~ 22.4mag. (23mag @ 2x106km)

Detecting ~1m @ 1.5 ~ 2.0 million km distance

• A 1m object at 1.5~2.0 million km is 22.4 ~23.0 mag

• (sphere 20% albedo Lambertian 0 phase angle)

• Sky background ~22.5mag/arcsec^2

• One low-cost example telescope(s) sensor to reach this faint limiting mag

• Quad 35cm telescope, 7sqdeg FOV, SNR~7.5 in ~15 min integration

• All 4 telescopes pointed at the same FOV (4 cameras sync’d)

• Cislunar space covers ~ 750sqdeg.

• Why a quad telescope?

• Lower cost than a single telescope with 2X the diameter (and 2X shorter length)

• Using 4 co-pointed telescopes vs 4 telescopes with adjacent FOVs but integrating 4X longer, potentially reduces computation costs, for many TBD algorithms.

770,000km

1.5x106km

Data sets for TBD2 Performers

• Real data sets

• Single telescope and quad telescope data sets.

• Actual noise from detectors

• Will contain detector defects, natural background that might cause false positives. (cosmic ray events, dense star field near galactic center, globular clusters, galaxies, etc.)

• Real data sets with synthetic RSOs (usually multiple synthetic RSOs to reduce the # of data sets to analyze)

• Quad telescopes with small but real differences in the 4 telescopes and their alignment.

• Synthetic data sets. (from LLNL)

• High fidelity simulations, with stars from the GAIA catalog. Sky background that represents a space-based telescope. RSOs will be accurately streaked, according to single image exposure time.

• RSOs, in different cislunar orbits simulated.

• LLNL and JPL will use a consistent set of physics assumptions in data generation

• Custom data sets available. We’re interested sensitivity vs integration time. If some TBD algorithms would produce higher sensitivity with a different single exposure time that the nominal, that will be accommodated.

What properties of different TBD algorithms are of interest

• Real data sets will have real detector artifacts and other properties that might affect the false positive rate.

• Synthetic data sets will provide observing conditions for a space based sensor.

• Sensitivity vs total integration time (for the baseline optical/camera system)

• Many TBD algorithms will establish a detection threshold at some value of SNR.

• At the detection threshold (because of photon noise of the RSO and sky background) the probability of detection is 50%.

• We’ll be interested in the prob of detection vs brightness of the RSO. (nominally set the sensitivity where the prob is 50%)

• False positive should be in the low 1-2% level. (more discussion next slide)

• Computational needs (and watts) to perform the analysis to keep up with real time.

• Would prefer < 300W, limit to < 600W.

• Different performers may want to use different GPUs/FPGAs, compute needs will be measured in watts, when running the algorithm

• If a custom compute accelerator is to be used the performer must provide a spare set of hardware to

LLNL.

• Astrometric accuracy is more important than photometric accuracy. For faint objects astrometric accuracy will be compared to what is possible relative theoretical limit given photon noise. We’ll also look at what the instrumental limits are when SNR is high.

SNR (signal to noise)

• Terminology and a more detailed scription of the technical terms

• SNR (signal to noise ratio)

• The signal is the #photons from the RSO.

• What is the noise? 1) sqrt(Nsky + RN^2) or 2) sqrt(Nsky +RN^2 + RSO)

• One has to define the term SNR more precisely

• 1) photometric SNR is the SNR of the photometric measurement (#2 equation)

• 2) detection SNR (use equation 1)

• What is detection SNR? Detection SNR is used to distinguish the presence of an RSO, versus the absence of the RSO. In the absence of the RSO a group of pixels can randomly appear above the mean causing a false detection. Where should the threshold be set?

• 1)where the threshold should be set is independent of what RSO may or may not be present.

• 2) often the threshold is set where the prob of a false positive is ~1%.

How Will Sensitivity be Evaluated

• Each TBD algorithm/software will be presented with data cubes with many synthetic RSOs of different brightnesses.

• We will interpret the detection threshold to be where the curve crosses 50%.

• For TBD algorithms that set the detection threshold at some value of SNR, 50% completeness will occur at the detection threshold.

• But not all TBD algorithms set their detection threshold at some given SNR.

• Often, the detection threshold is set where the false positive rate from just random gaussian (or photon) noise is 1%, 1 FP per 100 data cubes analyzed

20 20.5 21 21.5 22 22.5 23 23.5 24

Object Brightness magnitude

0.2

0.4

0.6

0.8

Pr ob ab ilit y of

D et ec tio n

Detection Probability vs. Object Brightness

Completeness curve for a system with Gaussian random noise and a threshold at 7.5σ

False Positive, Math speak vs English speak

• When someone says the false positive rate is 1%, the English meaning of this is that if we think we’ve detected 100 RSO’s one of them is not real.

• This interpretation does not make mathematical sense. (this interpretation is mathematical nonsense)

• In Math speak, 1% false positives means that if we look at 100 FOVs, on average we will have 1 false positive. (perhaps a better metric is # false positives per sqdeg of sky surveyed.)

• Why is the English interpretation nonsense?

• Let’s say we take a modern SSA telescope/camera/computer back to the year 1956, before the 1st RSO was launched. This SSA system might survey the sky for 1 night and if it has 1 false positive, that false positive rate is 100%.

• But if we take that same SSA system today and survey the sky and detect 1000 RSO with 1 false positive we have a false positive rate of 0.1%.

• The English interpretation of 1% false positives is NOT a measure of the SSA system. The 2nd interpretation IS a measure of the SSA system.

• False positive of 1% mean 1 FP per 100 data sets (on average).

• False positive can occur due to gaussian random noise (goal is < 1%)

• Artifacts in the camera, telescope, or the natural sky background. (also < 1%)