Zach Gazak, AFRL; Justin Fletcher, SMC/DirSP-G; Ryan Swindle, AFRL; Ian McQuaid, United States Air Force
Keywords: space situational awareness, deep learning, deep convolutional neural networks, computer vision, closely spaced objects
Abstract:
The detection of closely spaced artificial satellites informs critical tactical decisions for space domain awareness (SDA). One method, spectroastrometry, leverages differences in spectral energy distributions to detect the presence of multiple objects below the resolving power of a telescope. In this work we present a spectroastrometric approach leveraging convolutional neural networks (CNNs) to detect the presence of multiple objects in high resolution spectroscopic observations unresolvable through conventional imaging. Our proposed CNN is trained on a simulated dataset modeled after the SPICA II spectrograph. SPICA II–an optical instrument mounted on a trunnion port of the 3.6m AEOS telescope on Haleakala–is being used to build an on sky set of images to validate our approach. Using a dataset of 300,000 simulated observations, we train a CNN to perform classification between CSOs and single objects across distributions of on sky separations, satellite classes, and atmospheric parameters. Future work will utilize and observed spectral catalog to evaluate this approach with and without transfer learning and explore the applicability of this technique to alternate sensor designs. We quantify the theoretical limits of our approach across differences in CSO spectral energy distribution and relative and absolute target magnitude(s).
Date of Conference: September 15-18, 2020
Track: Machine Learning Applications of SSA