Konstantinos Tsaprailis, National Kapodistrian University of Athens, Greece & National Observatory of Athens; George Choumos, National Kapodistrian University of Athens, Greece & National Observatory of Athens; Charalampos Kontoes, National Observatory of Athens; Vaios Lappas, National Kapodistrian University of Athens, Greece
Keywords: Space Surveillance & Tracking, Artificial Intelligence, Initial Orbit Determination, Resident Space Object Characterization, Spin Status Estimation, Optical Observations, Tracking Data Message
Abstract:
The number of Resident Space Objects (RSOs) orbiting the Earth is continuously increasing, driven by the surge in satellite deployments and debris accumulation. This renders the sustainability of space operations a challenge, requiring effective Space Situational Awareness (SSA). In this study, we introduce AI-based approaches that enable the transition from single telescope observations to a characterized and catalogued space object, incorporating spin-state classification and initial orbit determination. Optical telescopes are the most commonly available type of ground-based sensor for space surveillance, considering not only their number, but also their distribution over the globe. We evaluate our AI methods on real optical observations, coming from i) the Mini-Mega Tortora (MMT) system, and ii) the “Kryoneri” telescope, a Greek operational sensor that participates in the European Union Space Surveillance & Tracking (EUSST) network.
The first part of the study presents a novel approach for predicting the spin status of space objects using a convolutional neural network (CNN) architecture applied to optical sensor observations. Our methodology leverages data from the MMT system, comprising lightcurves from 12,494 distinct objects, to train a deep learning model capable of classifying space objects as spinning or non-spinning from a single Tracking Data Message (TDM) of optical observations. The CNN architecture employs multiple convolutional layers to capture temporal patterns at different scales. We conducted rigorous validation using 5-fold cross-validation to ensure robustness. Results demonstrate consistent state of the art performance across all folds, with an average accuracy of 92.8% (ranging from 91% to 95%). The model exhibits high precision (98%) in identifying non-spinning objects and good recall for spinning objects (ranging from 90% to 94%), yielding weighted F1-scores between 0.92 and 0.95 across all folds. The performance metrics indicate that our approach can reliably classify spin states from limited observational data and we consider this result to be state of the art in spin status prediction using the data from the MMT database.
The second part of the study presents an AI approach to perform Initial Orbit Determination (IOD) from single telescope observations. Traditional methods for IOD, such as Gauss and Gooding, rely on a small number of observations and struggle with accuracy, especially on short observational arcs. While the orbital plane determination can be relatively accurate, the lack of range information leads to weakly constrained along-track motion. Our AI-approach for IOD comprises a set of independent Random Forest Regressors, each individually tailored to predict a single orbital element. The tailoring involves the definition and combination of the most informative features for each of the orbital elements, and the hyperparameter tuning of the models. Our IOD dataset includes > 7000 real TDMs with right ascension and declination angles, collected from observation campaigns of the Kryoneri tracking telescope between September 2023 and July 2025. To train the models, we extract our ground-truth orbital elements from the most recently available TLE file prior to each corresponding observation night. Our results showcase significant reduction in the Mean Absolute Error (MAE) of all orbital elements compared to the Gauss traditional method, namely ~99.5% in eccentricity, ~98% in semimajor axis, ~95% in inclination, ~19.5% in the argument of perigee and ~37% in the right ascension of the ascending node. In addition, our AI approach properly handles cases where traditional methods lead to highly uncertain or hyperbolic orbits, a common occurrence on TDMs with limited number of observations.
These methodologies directly support activities within the Greek National Operations Center for SST, relevant to the processing of uncorrelated TDMs and to the scheduling of follow-up observations, thus optimizing sensor tasking strategies.
This research demonstrates that AI can bridge the gap between traditional methods and modern SSA needs. By supporting the move from single real optical observations to cataloged and characterized space objects, we enhance the efficiency, accuracy, and scalability of space object tracking and cataloging, ultimately contributing to the long-term safety, security, and sustainability of space operations.
Date of Conference: September 16-19, 2025
Track: Satellite Characterization