Machine Learning Approaches for Sub-Catalogue Space Debris Clusters Identification in High-altitude Orbital Regions

The growing population of space debris in high-altitude orbital regions, such as GEO and GTO, poses increasing risks to the safe operability of strategic assets. Identifying clusters of objects is essential to better characterize the orbital environment at these altitudes and to support statistical models such as the European Space Agency’s (ESA) MASTER tool. The Astronomical Institute of the University of Bern has significantly contributed to this effort by conducting, on behalf of ESA, long-term optical surveys for sub-catalogue debris in high-altitude orbits. Given the large and complex datasets involved, machine learning (ML) techniques offer a promising approach to reveal physically meaningful patterns. This paper investigates the application of ML-based clustering techniques to identify space debris clusters in high- altitude orbital regimes. Two datasets are analysed: a simulated dataset generated with ESA’s PROOF tool, based on the MASTER environment model, and an observational dataset collected with ESA’s 1-meter telescope at the Optical Ground Station in Tenerife. Several clustering techniques are first applied to the simulated dataset using inclination and Right Ascension of Ascending Node as features. Density-based algorithms, in particular DBSCAN, successfully identify meaningful clusters, including uncontrolled GEO objects following the solar–lunar perturbation-driven evolution path in the (i–Ω) plane. The analyses are then extended to higher-dimensional datasets, where DBSCAN distinguishes GEO and GTO populations and further resolves subclusters within the GTO regime, likely linked to breakup events. The same analyses are applied to the observational dataset where the GEO evolution path is clearly detected, along with minor clusters in the inclination range [2.5°, 12.5°] consistent with GTO debris. In higher dimensions, DBSCAN identifies a main GEO cluster. However, because orbit determination assumes circular orbits, eccentricity information is unavailable, preventing a clear GEO/GTO separation. Nevertheless, the mean motion distribution strongly suggests an association with GTO orbits.