A recurrent neural network-based approach for ballistic coefficient estimation of resident space objects in low earth orbit

Characterizing Resident Space Objects (RSOs) has become paramount for several Space Situational Awareness functions, such as accurate orbit prediction, collision avoidance and sensor tasking. Due to the huge and diverse amount of data to handle and fuse for this purpose, there is an increasing interest in Machine Learning (ML) approaches to retrieve physical parameters of RSOs in a costefficient manner. Among different ML architectures, Recurrent Neural Networks (RNNs) are particularly suitable to handle sequential or time-series data. In this context, this paper demonstrates the possibility to use RNNs to estimate the ballistic coefficient of RSOs in Low Earth Orbit from time series of orbital elements. A particular RNN architecture has been adopted, i.e., the Gated Recurrent Unit, as it has been demonstrated to be cost efficient and to have good performance for the required application. The sensitivity of RNNs with respect to atmospheric model uncertainty has been analysed. The algorithm can handle unevenly spaced time-series data. The proposed neural network has been demonstrated to be robust with respect to orbital state errors. The applicability of the presented approach is tested and discussed using synthetic datasets generated with a high-accuracy numerical orbital propagator, reaching a mean percentage error of 10% in the estimation of the ballistic coefficient. (c) 2024 COSPAR. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).