Rithwik Neelakantan, Digantara; Mridul Songara, Digantara; Latha S, Digantara; Tanveer Ahmed, Digantara; Anirudh Sharma, Digantara Industries
https://doi.org/10.64861/MDJS3372
Keywords: Satellite Maneuver Pattern of Life, Maneuver Prediction, Battlespace Characterisation
Abstract:
Gen. Saltzman’s theory of Competitive Endurance suggests that U.S. Space Forces avoid operational surprise, deny adversaries first-mover advantage, and campaign responsibly across the competition continuum. Each tenet hinges on a foundational question that current Space Domain Awareness (SDA) capabilities only partially answer: not just what an adversary satellite has done, but what it will do next. This paper addresses that gap directly, proposing a predictive maneuver framework based on the historical pattern of life, designed to give warfighters anticipatory insight into non-cooperative and adversarial Resident Space Object (RSO) behavior.
The third paper in this series builds on a previously established multi-modal maneuver detection and Pattern of Life characterization pipeline and introduces a third predictive layer that converts historical behavioral signatures into operationally actionable forecasts. The first two papers of the series involved: 1) Maneuver detection using Two-Line Element (TLE) analysis via an iterative sequential window method & Hidden Markov Model-based state estimation and 2) Generation of a historical baseline for maneuver and quantification of deviation of pattern from that baseline. Specifically, the baseline was generated by characterising maneuvers across a ten-dimensional feature space encoding post-maneuver orbital elements, delta-V estimates, and inter-maneuver intervals and then DBSCAN orbital clustering and Mahalanobis distance based PoL-violation scoring from the baseline.
On this foundation four complementary single-object models for SJ-25 (NORAD ID 62485) are built. Each targets a distinct operational question – when the next maneuver will occur, whether one maneuver triggers others, which behavioral regime the object currently occupies, and what type a given maneuver is – so the four are complementary rather than competing, and none is redundant with another. These four models are: 1) a Bayesian renewal process for maneuver timing 2) a two-regime Hawkes process for self-excitation 3) a hidden semi-Markov model for regime identification and 4) a Gaussian-process classifier for maneuver type with calibrated uncertainty. A generalization scheme follows, extending timing prediction to other satellites via a three-level hierarchical Bayesian renewal model, whose partial pooling forecasts even sparsely-observed objects that per-satellite methods cannot address at all. The data used for the analysis is gathered from various sources including JCO-issued NOTSO (Notice to Space Operators) messages and in-house-generated maneuver detections, spanning the SJ-25 record from September 2025 to June 2026.
A key methodological finding emerging from this work is the importance of behavioural regime separation prior to prediction. When Rendezvous and Proximity Operations (RPO) maneuvers are included in the training data alongside routine station-keeping events, the high variability of RPO-driven maneuver timing contaminates the prediction signal and inflates apparent model uncertainty. Restricting analysis to non-RPO maneuvers – those reflecting purposeful orbital maintenance behaviour rather than proximity approach profiles – yields a qualitatively distinct and more predictable pattern. This behavioural regime-separation finding recurs independently across all four single-object models. It is operationally significant, suggesting that distinct prediction models should be maintained for station-keeping and proximity-operations phases.
Forecasts are evaluated out-of-sample using a walk-forward protocol, scored by the continuous ranked probability score (CRPS) against climatology and persistence baselines. The hierarchical model is well-calibrated – its 90% prediction intervals cover 89–92% of outcomes – but does not improve on a regime-conditioned climatology in point accuracy (23.6 vs. 23.0 days CRPS), because geostationary station-keeping intervals are inherently highly variable. Its real value therefore lies in calibrated, regime-aware probabilistic forecasts and in covering barely-observed satellites that per-satellite methods cannot forecast at all, rather than in raw accuracy. Negative results include the statistical insignificance of self-excitation and the failure of physical covariates to improve skill. The framework is also cross-validated against SJ-25’s RPO partner, SJ-21 (NORAD ID 49330): an independent detector, extended from a single orbital element to the full element set, recovers all three catalogued station-keeping burns (versus one before) and surfaces a previously unlogged inclination-plane change.
The operational output is a ranked, time-tagged prediction of probable adversary maneuver activity delivered to analysts before the maneuver occurs – transforming SDA from a reactive cataloging function into a proactive warfighting tool. When a warfighter can anticipate a maneuver, sensor tasking can be pre-positioned, conjunction screening initiated proactively, and intent assessments updated before the adversary acts. The methodology is orbital-regime agnostic, validated across LEO and GEO, and designed for integration into existing SDA decision-support pipelines with provisions for real-time recalibration as new observations are ingested.
Date of Conference: September 15-18, 2026
Track: Satellite Characterization