David Beskow, United States Military Academy; Christopher Gramling, United States Military Academy; Madison Oliver, United States Military Academy
https://doi.org/10.64861/ONHU1726
Keywords: Space Domain Awareness, Satellite Behavior Characterization, Geosynchronous Earth Orbit, Behavioral Embeddings, Maneuver Detection, Long Short Term Memory, Representation Learning, Pattern of Life, Behavioral Clustering
Abstract:
Space Domain Awareness increasingly requires more than binary maneuver detection: analysts need
compact, comparable descriptions of how individual geosynchronous (GEO) satellites behave relative to the rest of the constellation. We show that the hidden state of a supervised Long Short-Term Memory (LSTM) maneuver detector, trained on physics-informed Radial-In-Track-Cross-Track (RIC) and topocentric residuals from electro-optical observations, can be repurposed as a satellite behavioral embedding without any additional retraining. Track-level hidden states are aggregated into one 128-dimensional descriptor per satellite via a probability-weighted mean that emphasizes maneuver-relevant tracks without a hard detection threshold. Applied across all 65 satellites in our GEO dataset, the resulting embedding space organizes satellites along an axis correlated with overall maneuver activity and detection confidence, recovers spatially coherent clusters under both residual frames, and exhibits a more internally consistent similarity structure under the physics-informed RIC frame than under the topocentric baseline. These results indicate that a representation learned purely for supervised detection retains structure relevant to broader pattern-of-life characterization. Detection performance for this classifier is reported later in this paper; broader comparisons across residual frames and satellite counts are outside its scope.
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications