Gregory Sallaberry, Lawrence Livermore National Laboratory; Robert Armstrong, Lawrence Livermore National Laboratory; Nathan Golovich, Lawrence Livermore National Laboratory; Alex Geringer-Sameth, Lawrence Livermore National Laboratory; Luc Peterson, Lawrence Livermore National Laboratory; Rishay Jain, University of California Berkeley & Lawrence Livermore National Laboratory; Ziad Ramadan, University of California Berkeley & Lawrence Livermore National Laboratory
https://doi.org/10.64861/WVXH6624
Keywords: Astronomy, Computing, Big Data, Survey, Cislunar, Ground-Based
Abstract:
In the current space environment where cislunar activity is steadily increasing, it is important that we can detect and characterize faint objects for space situational awareness. Wide-field astronomical surveys and telescopes, including the Zwicky Transient Facility (ZTF) and the upcoming Legacy Survey of Space and Time (LSST) being conducted at the Vera C. Rubin Observatory, will collect an unprecedented amount of data (e.g., on the order of 20TB of imaging for LSST per night) and have the ability to detect small and distant objects by virtue of their large collecting areas (8.4 m primary mirror for LSST) and fields of view (9.6 sq. deg for LSST and 47 sq. deg for ZTF). While the survey designs are motivated by science cases, they will necessarily capture large numbers of objects moving in the Earth-Moon system, which will appear as streaks on images. We are developing a pipeline to identify and characterize these streaks, fit orbits to each detected object (with quantified uncertainties), and upload the results to a central database within minutes of image capture.
We use the LSST pipeline for calibration and reduction to construct difference images for analysis. Difference imaging helps mitigate false positive streak detections by reducing the influence of stars and galaxies in exposures. Using the Fast Radon Transform (Nir, 2018), we identify streaks in an image by convolving the point spread function with all possible line segments that run through the image, enabling us to detect faint streaks below the single-pixel noise level. We expect to reach a depth of 21.5 magnitudes for a typical GEO object. The advantage of using the Fast Radon Transform method is a reduction in computational complexity from N3 to N2log(N), where N is the length of the image grid in pixels.
Each detected streak is characterized by the full 4×4 covariance matrix on its apparent sky position and transverse motion. The measurements and covariance are then used in a Bayesian inference step to predict the object’s three-dimensional position and velocity vectors at the time of observation. The two unmeasured components of the state vector (geocentric distance and its rate of change) are controlled with a data-and simulation-driven prior distribution that is conditional on the observable components. The result is a posterior sample of state vectors, which can be summarized by a central value and 6×6 covariance. These state vectors can be correlated to known objects and results will be sent (both state vector inference and the streak measurements) to the Unified Data Library within minutes of each observation, allowing for rapid, targeted follow up on any single detection where relevant.
Since our analysis is agnostic to the identity of the objects we observe, this allows us to determine streaks that may belong to unknown objects, which would make them a priority for follow-up. During daytime hours, when the pipeline is not actively being run to detect new objects, we will use our compute resources to propagate the orbits of all known objects to a set of times the following night. This propagation allows us to correlate the following night’s set of observations faster, as it will be an interpolation rather than propagation. This is particularly powerful for slow-moving objects in cislunar orbits that will be seen multiple times per night by survey telescopes.
Date of Conference: September 15-18, 2026
Track: SDA Systems & Instrumentation