Zach Gazak, Space Systems Command A&AS; Ryan Swindle, Space Systems Command A&AS; Sierra Morales, University of Hawaii at Manoa; Matthew Phelps, Space Systems Command A&AS; Kevin Iott, PlaneWave Instruments; Eric Blackhurst, Planewave Instruments; Justin Fletcher, Space Systems Command A&AS
Keywords: star detection, neural networks, robotic observatories, deep learning, machine learning
Abstract:
The proliferation of increasingly mobile assets in Earth orbit — and the expansion of organizations in control of those assets — demands a proliferation of ground based space domain awareness (SDA) infrastructure to maintain safe access to — and operations in — space. Critically, this includes mass-produced 0.5-1.5 meter optical telescopes which are rapidly deployable, ready for use within days of deployment, and provide precise metric observations and enhanced data products without continued human intervention. On the hardware side, this is largely solved by eschewing custom development and embracing commercial telescopes which are already optimized for cost and quality. In this work we focus on the software and analysis requirements for initial system calibration and ongoing high-precision star detection enabling the generation of metric observations required for orbit determination and tracking. To this end we proceed in three parts. First, we demonstrate a toolset (“Burr”) for automatic first light calibration of new telescopes, including the generation of real, diverse star streak datasets. Second, we describe a new concept of operations using both rate and sidereal tracking and an analysis suite (“SENPAI”) which returns ~1 arsecond astrometric residuals on calibration satellites in right ascension and declination. Finally, we introduce a custom lightweight neural network (“StarCSP”), which borrows from the crowd counting literature to avoid pitfalls of learned star detection. In combination, this set of tools (Burr, SENPAI, and StarCSP) provide the necessary framework to convert an optical telescope into a high precision metric observation asset within days of deployment. We demonstrate this on 0.35, 0.5, and 1 meter systems.
Date of Conference: September 16-19, 2025
Track: Machine Learning for SDA Applications