Miguel Velez-Reyes, University of Texas El Paso; Dan DeBlasio, Carnegie Mellon University; Luis Cedillo, University of Connecticut; Aryzbe Najera, University of Texas El Paso; Nicholas Duran, University of Texas El Paso; George von Borries, Universidade de Brasília
https://doi.org/10.64861/YYKM6386
Keywords: spectral signature analysis, machine learning models, unresolved imaging, satellite identification, simulation models, spectrophotometry,
Abstract:
The increasing congestion of the space domain creates a critical need for improved techniques to identify and characterize Resident Space Objects (RSOs). Networks of small-aperture telescopes (≤ 1 m) provide a cost-effective means for persistent monitoring and can be readily deployed at multiple geographic locations to increase coverage and revisit rates. However, many RSOs observed from ground-based small-aperture telescopes cannot be spatially resolved due to small size (e.g., nanosatellites or orbital debris) or large orbital distances (e.g., GEO, MEO, or cislunar). Consequently, information extraction must rely on the properties of the reflected light rather than spatial imagery. Hyperspectral sensing and slitless spectroscopy provide rich spectral (and potentially spectro-temporal) signatures that contain information related to spacecraft materials, geometry, attitude, and operational state.
This paper presents an overview of machine learning experiments conducted within the Science of Non-Resolved Space Object Signatures for Space Domain Awareness (SoS4SDA) project using simulated slitless spectrograph imagery of unresolved resident space objects (URSOs). Two deep learning architectures, TriFunNet and TriFunNet-KL, were evaluated for satellite identification from unresolved hyperspectral imagery. Both architectures achieved strong classification performance of approximately 82% recall and precision. However, TriFunNet-KL produced a significantly more structured latent discriminative embedding (LDE).
To evaluate the utility of this embedding for SDA applications, a series of leave-one-satellite-out experiments was performed in which one satellite class was treated as an unknown URSO while the network was trained on the remaining data. Results showed that observations from previously unseen satellites were consistently mapped to distinct regions of the latent space rather than being merged with known object classes. These findings suggest that learned latent representations from hyperspectral observations can support not only satellite identification but also other SDA objectives.
Date of Conference: September 15-18, 2026
Track: Satellite Characterization