Model validity dynamics in learned spectroscopic recognition of satellites

Abstract

Recent work demonstrates that convolutional neural networks can be trained to recognize artificial satellites from spatially unresolved ground-based observations (SpectraNet). SpectraNet enables space domain awareness (SDA) catalogs to be enriched with object identity, a critical source of information for space domain stakeholders. As learned spectral SDA matures, conditions for training and deploying performant and calibrated neural network recognition algorithms must be measured. In this work we present a simulated three year baseline of observations using a longslit spectrograph on a single telescope. We use this dataset to develop a framework for measuring baseline data requirements for performant SpectraNet models, and for testing the performance of those models after deployment. On this limited (single telescope, longslit spectrograph) setup, the presented framework returns a performant model after three weeks of collections. Further, we find that a model can be deployed for a full annual cycle after twenty six weeks of data collection, and the model reaches maximum sustained inference performance after a year. Thus a SpectraNet powered longslit spectrograph can provide tactical inferences after a few weeks and be retrained to infer through seasonal variability during deployment. We find that the simulated system and dataset regularly exceed 82% classification accuracy, and discuss performance improvements with enhanced instrumentation and/or multi-telescope networks.