ABSTRACT Significant research in the field of space domain awareness (SDA) has focused on improving AI‐driven data processing and classification tasks. Previous studies have explored the classification of orbiting man‐made object types such as satellites, rocket bodies, and debris, yet there is a noticeable gap in the literature concerning the subclassification of debris shapes such as fragments and detached satellite components. This lack of focus on debris characterisation despite the growing urgency to study Earth‐orbiting debris could be attributed to the scarcity of labelled debris data. More importantly, debris shape plays a crucial role in collision risk assessment, reentry prediction, and active debris removal (ADR). In the absence of publicly available datasets with detailed shape information, this study establishes a baseline for debris sub‐classification, aiding in improved debris mitigation and collision avoidance efforts. To address these challenges, a light curve simulation framework was created to generate LEO debris light curves based on physical object parameters and initial conditions defined by historical two‐line elements (TLEs) of debris. The principal investigation involved debris shape classification using a long short‐term memory fully convolutional network (LSTM‐FCN). An ablation study was carried out to investigate the performance of the LSTM and FCN separately. In addition to debris shape, the light curves demonstrated a level of sensitivity to material type. This motivated a secondary study involving multi‐task learning (MTL), in which material classification was introduced to the original LSTM‐FCN. The results demonstrated that the MTL approach enhanced the model's generalisation for the shape classification task. A 2% improvement from the single‐task to the multi‐task model is considered notable, highlighting the benefits of MTL. Retrieving material and shape information indirectly informs classification tasks in SDA on the debris' sensitivity to both atmospheric drag and solar radiation pressure, which are key considerations in the study of debris motion and ADR. Future work will focus on incorporating irregular shapes into the dataset and exploring the impact of a larger dataset on classification performance.
Adriano et al. (Mon,) studied this question.