A Hybrid Deep Learning Framework for Space Debris and Orbital Decay Prediction Using DISCOS

Space debris consists of defunct objects in orbit that pose collision risks to satellites and spacecraft, with long lifespans and growing density threatening space sustainability. Predicting their orbital decay is crucial but challenging due to atmospheric fluctuations, incomplete data on the objects, and noisy tracking measurements. Existing detection and decay prediction methods combine radar, optical tracking, physics-based models, and deep learning approaches (e.g., YOLO and CNN hybrids), but face limitations in accuracy and reliability, especially for small debris and uncertain conditions. To address these challenges, we utilized the DISCOS dataset, which provides detailed records of space objects and their decay dates, to investigate patterns of orbital decay. Our study introduces a novel combination of autoencoders and multilayer perceptrons (MLPs), which enables effective feature extraction and achieves 93 % accuracy in classifying space debris. We also applied XGBoost and UMAP with DBSCAN for unsupervised clustering of decay patterns. Together, these methods improved the accuracy of decay prediction and classification of debris reentry events.