Enrique De Alba, EO Solutions; Kimmy De Alba, USSF SSC A&AS; Zach Gazak, USSF SSC A&AS; Justin Fletcher, USSF SC A&AS
https://doi.org/10.64861/EFJB2607
Keywords: Space Domain Awareness, Electro-Optical Imaging, Tracked Satellite Imagery, Anomaly Detection, Component Classification, Satellite Damage Detection
Abstract:
Detecting structural change in resolved imagery is one component of ground-based optical characterization of resident space objects (RSOs), where atmospheric turbulence degrades every frame. It remains unclear how much apparent change-detection performance depends on references operational sensors cannot supply. We splice frame-aligned intact and component-removed International Space Station (ISS) renders at known transitions, yielding deterministic per-frame labels and no-change controls. We evaluate one learned relational system under increasingly realistic reference protocols, each audited by scoring no-change controls as events. Training detectors to ignore viewing geometry discards the structural evidence, collapsing localization (posterior mass within $pm 10$ frames of the true onset) from 0.774 to 0.068 on held-out poses. Reference comparison avoids this penalty, but localization reaches 0.858 against each frame’s intact twin and only 0.416 against a rolling reference blind to the change time. The audit rejects a reference policy frozen at the change time: it localizes no-change splices (0.965) as readily as events (0.943). A translation-equivariant convolutional localizer improves blind-reference localization to 0.486 with a one-frame median onset error. Sequence-level detection remains unresolved, and false-alarm calibration stays local to viewing geometry. More broadly, our work positions the reference protocol as an essential, reportable component of change-detection claims under degraded imaging.
Date of Conference: September 15-18, 2026
Track: Satellite Characterization