George Landon, Cedarville University; Timothy Giblin, i2 Strategic Services LLC; Kyle Jones, United States Air Force Academy; Francis Chun, United States Air Force Academy
https://doi.org/10.64861/NFWB6289
Keywords: Geostationary Earth Orbit, Unresolved satellite characterization, Slitless spectroscopy, Domain-adversarial training, Feature alignment, 1D convolutional neural networks, multi-head classification
Abstract:
Increasing satellite density in Geostationary Earth Orbit (GEO) requires cost-effective observation networks to maintain Space Domain Awareness (SDA). While spatially resolved imaging of GEO satellites requires large-aperture telescopes, distributed commercial-off-the-shelf optical sensor networks, such as the Falcon Telescope Network, equipped with slitless spectrographs provide an accessible alternative. However, extracting spacecraft characterization from unresolved 1D spectral signatures is an underdetermined inverse problem, complicated by atmospheric extinction, solar illumination geometry, and cross-site instrumental disparities.
This work presents a multi-task spectral classification framework coupled with a decision-support interface. The method uses a 1D ConvNeXtV2 backbone integrated with Feature-wise Linear Modulation (FiLM) metadata fusion, Discrete Cosine Transform (DCT) continuum augmentations, and Wasserstein Domain-Adversarial Neural Network (W-DANN) critics to reduce site-dependent and temporal variations. To address spectral ambiguity, the taxonomy is structured as a Directed Acyclic Graph (DAG) with residual Adapters, predicting upstream attributes (Total Solar Panels, Radio Frequency Bands, Bus Family) alongside downstream Bus Variant and Platform identity.
Trained under multi-site domain variation across historical collections (2022 through December 11, 2025 across USAFA-16, USAFA-Falcon, and OJC) and evaluated on a time-forward holdout set (December 30-31, 2025 at OJC) to assess temporal generalization after learning from cross-site heterogeneous data, the network recovers at least one correct hierarchical prediction in 90.7% (88/97) of holdout observations (?2 correct in 71.1%, ?3 in 62.9%, ?4 in 49.5%, and ?5 in 38.1%).
Task success rates reach 76.3% for RF Bands (measured as at least one top-predicted band match), 62.9% for Total Panels, 60.8% for Bus Family, 53.6% for Bus Variant, and 58.8% for Platform identity. The trained pipeline is integrated into a decision-support interface that presents ranked probabilistic hypotheses under temporal and environmental variability.
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications