Improving ballistic coefficient estimation of resident space objects in low earth orbit

Characterizing Resident Space Objects (RSO) is fundamental for several Space Situational Awareness functions, such as accurate orbit prediction and collision avoidance. The main non-conservative orbital perturbation that acts on RSOs in Low Earth Orbit (LEO) is the atmospheric drag, that can be predicted with significant uncertainty. In fact, it strongly depends on the atmospheric density, which has a stochastic nature also due to the effect of the solar activity, and on some physical characteristics of the RSO, which are either unknown or known with limited accuracy for space debris. The ensemble of physical characteristics affecting the drag-based acceleration experienced by an RSO is represented by the ballistic coefficient (BC), i.e., the product between the drag coefficient (Cd) and the area-to-mass ratio (AMR). Current literature proposes several approaches to estimate either the single terms that constitute the BC or the whole parameter. In particular, many recent works are investigating Machine Learning (ML) and data driven techniques, due to their advantage in terms of computational efficiency. Most of these works, either assume that the ballistic coefficient is constant over the time interval of the analysis or estimate a mean value. However, the BC of a space object can vary significantly over a certain time interval, especially for space objects for which the combination of shape and attitude dynamics determine a significant time variation of their cross section, which has a non-negligible impact on the orbital trajectory. In this context, this paper provides two main contributions. First, it proposes a reduced set of features to train Neural Networks for ML-based ballistic coefficient estimation of RSOs, while also investigating two preprocessing techniques to reduce the computational cost of the training phase and to filter the astrometric data received in input by the trained network, thus removing undesired oscillations. Secondly, the potential of a ML-based approach to detect ballistic coefficient variation for space objects in LEO is investigated. A sensitivity analysis is conducted to assess the minimum variation of the BC that can be detected for a fixed time interval, as well as the minimum time needed to detect a certain BC variation. The applicability of the presented approaches is tested and discussed using synthetic datasets.