Zach Gazak, SSC/SZG; Ryan Swindle, USSF/SSC; Matthew Phelps, USSF SSC/SZG; Justin Fletcher, USSF SSC/SZG
Keywords: Spectroscopy, Machine Learning, Neural Networks, Small Telescopes, Hyperspectral
Abstract:
The protection of high value government and commercial satellites is predicated on the detection, recognition, and localization of active and defunct satellites, rocket bodies, and debris across all orbital regimes above Earth. In the bulk of those orbital regimes, recognition has lagged the needs of space domain awareness stakeholders, as target distances preclude resolved imagery from ground-based observatories. Recent research has demonstrated that the rich, distance-invariant information encoded into the spectrum of an object’s reflected sunlight can fill this need. This concept, named SpectraNet, introduced deep learned spectroscopy as a solution bypassing the dearth of exquisite knowledge of the physics governing reflection against complex materials. By training deep neural networks against a baseline of observations, SpectraNet learns the information content linking spectra to object identity. In this work we introduce and demonstrate a significant advancement in the SpectraNet pedigree which solves a number of limitations with the original proof of concept. By utilizing multi-spectral imaging sensors instead of a traditional longslit spectrograph, this work eliminates the need for a priori knowledge of target location. Further, targets are simultaneously detected and identified, and the stellar background allows for the extraction of astrometric information required for target localization. Finally, this technology enables proliferation by utilizing inexpensive, off the shelf components and small telescopes. In this work we describe the learned backbones for detection and recognition, end to end performance in simulation, and initial on sky observations.
Date of Conference: September 19-22, 2023
Track: Satellite Characterization