Using Machine Learning Methods for Tracklet Association
Please login to view abstract download link
The amount of space debris in the Earth’s orbit presents a significant threat to satellites and other operational spacecraft as even smallest objects can cause significant damage upon impact. To keep track of controlled and uncontrolled orbital objects the BACARDI-Software (Backbone Catalogue of Relational Debris Information) was developed by the German Aerospace Center. In a fundamental aspect of this Software observations of space objects, such as tracklets, are assigned to either known orbits of established orbital objects or to other unassigned tracklets. The complex algorithms of Tracklet-Object-Correlation and Tracklet-Tracklet-Correlation are known to be time-consuming and in the latter case not always reliable [1]. It was already shown that graph-based clustering methods could reduce the amount of false associations in the Tracklet-Tracklet Correlation, while also improving the computational speed [2]. To further enhance the process of tracklet association, decision-tree-based classification algorithms of Random Forest and XGBoost are used to improve accuracy, reliability and computational efficiency. Both approaches yield similar results and accuracies on the test data set and can significantly increase the number of correct associations, which would reduce the risk of collisions of tracked space debris.
