Machine Learning-Based Classification of Space Debris and Satellites Using Orbital Parameters

The classification of space debris and active satellites is crucial for maintaining space situational awareness and preventing collisions. This study presents a machine learning approach that utilizes two-line element (TLE) data to distinguish between satellites and debris based on key orbital parameters, including eccentricity, inclination, mean motion, right ascension of the ascending node (RAAN), and semi-major axis. Two modelsa feedforward neural network (FNN) and a random forest classifier- were developed and evaluated for their classification performance. The TLE-derived parameters were preprocessed to address data inconsistencies, with the feature importance analysis revealing the inclination and the semi-major axis as dominant discriminators. The Random Forest model utilized an ensemble of decision trees with Gini impurity criteria, while the FNN employed a three-layer architecture with ReLU activation and dropout regularization. The random forest classifier achieved superior accuracy (92.4%) compared to the FNN (87.1%), demonstrating robustness to noisy TLE data and non-linear feature relationships. Both models significantly outperformed traditional clustering-based taxonomic methods, particularly in handling edge cases such as debris with satellite-like orbits. This data-driven approach provides a scalable alternative to resource-intensive optical or radar-based classification systems. Future work will integrate the temporal evolution of orbital parameters to address decaying orbits and the detection of maneuvers.