J. Zachary Gazak, United States Space Force; Ian McQuaid, Air Force Research Laboratory; Brandon Wolfson, KBR; Justin Fletcher, United States Space Force
Keywords: Space Domain Awareness, Spetroscopic positive identification, spectroscopy, convolutional neural networks, incremental learning, learned positive identification
Abstract:
Positive identification of resident space objects (RSOs) injects rich information to traditionally position-only space traffic catalogs. Prior simulation work has demonstrated that convolutional neural network classifiers trained on spectroscopic data (SpectraNets) are highly effective, with accuracies on static 64 class problems exceeding 90%. In practice, a deployed SpectraNet must respond to a dynamic dataset, in which the number of observations of known RSOs grow and new RSO classes are periodically added to the problem. We present a paradigm in which model retraining with incremental learning improves deployed effectiveness of SpectraNet and capitalizes on the natural increase in size and diversity of the underlying dataset afforded by autonomous collection. This work provides a pathway for the deployment of SpectraNet-class models to live telescope operations. We demonstrate the dynamics of SpectraNet with the introduction of new RSOs, determine that a minimum of 100 observations is needed to reach performance on a new class of RSO, and show how a subset of incremental learning techniques are well suited for spectroscopic imagery.
Date of Conference: September 14-17, 2021
Track: Machine Learning for SSA Applications