Chad Mello, United States Air Force Academy; David Strong, Strong EO Imaging, Inc.; Kyle Jones, Ardent Technologies, Inc.; Timothy Giblin, i2 Strategic Services LLC; Francis Chun, United States Air Force Academy
https://doi.org/10.64861/TMGI2270
Keywords: GEO identification, space situational awareness, SDA, slitless spectroscopy, unresolved satellite spectra, Falcon Telescope Network, deep learning, generative augmentation, edge AI, edge compute, near-real-time, on-site processing, latency and throughput
Abstract:
Building on our previous work classifying geosynchronous orbit (GEO) satellites from slitless, non-resolved optical spectra, this study expands both the dataset and the modeling workflow to improve robustness and operational readiness. Attempting to scale our original deep-learning classifier to a broader observation-level corpus revealed unexpected limitations, triggering a comprehensive audit of both the legacy evaluation protocols and the dataset representation. We demonstrate that previous high-accuracy estimates reflected untested boundary conditions within the augmentation protocol rather than strict cross-night generalization, prompting us to establish a rigorous “night-disjoint” baseline that better reflects operational environments.
This work is designed to integrate directly with a newly implemented automated Falcon Telescope Network (FTN) pipeline that is beginning to deliver fresh, standardized spectral products. By evaluating deep architectures against a full suite of classical statistical methods, we find that simple correlation-based models (e.g., 1-Nearest Neighbor and Extra-Trees) currently outperform deep networks on our limited sample sizes and are highly efficient. These classical models are therefore recommended to operate as the initial edge-deployed processes at the sensor location, providing rapid, confidence-scored classification and near-real-time generation of machine-consumable products for downstream exploitation.
Rather than relying solely on synthetic augmentation to mitigate data scarcity, this study identifies the physical and processing bottlenecks — such as near-spectral-twin ambiguity and night-level collection signatures — that limit current model performance. We provide concrete collection recommendations, including targeting 10-20 observations per night to make physically motivated solar-analog scaling viable. Collectively, this work establishes a rigorous, defensible baseline for optical GEO satellite identification. Furthermore, we maintain our original hypothesis that deep models will ultimately outperform classical approaches; this study charts the explicit operational path for consistent, automated FTN data collection required to realize that deep-learning scalability over time.
Date of Conference: September 15-18, 2026
Track: Satellite Characterization