Astrometric Data for Early Time Series Classification of Resident Space Objects

The classification of resident space objects (RSOs) is crucial for space situational awareness due to its importance in national defense and the rapid growth of commercially launched RSOs. Machine learning (ML) and deep learning (DL) techniques now facilitate classification of RSOs based on sensor observations; however, they are often constrained by the lack of a sufficiently large training dataset. This work leverages supervised ML algorithms and DL techniques for early time series classification of RSOs, using the first large-scale and long-time RSO coverage dataset, consisting of vector covariance messages (VCMs) covering 22,303 RSOs over six months. Early classification of RSOs is of key interest to space system operators; therefore, the acceptable tradeoff between accuracy and timeliness of the prediction is investigated. For the first time, a transformer model is applied to the classification of RSOs using irregularly sampled astrometric time series, such as position and velocity measurements. Despite the variability in RSO’s VCM estimates and class imbalance, this approach achieved the best classification results among all the techniques considered, with F1 scores (the harmonic mean of precision and recall) above 0.80 for all RSO classes using as few as 5–15 measurements. §§