A Privileged Information-Based Framework for Satellite Micro-Motion Recognition

As human space activities rapidly expand, space situational awareness and behavior understanding of non-cooperative space targets have become critical for orbital security. Radar micro-Doppler (m-D) signatures provide valuable insights into the fine-grained micro-motions of space targets. However, under low signal-to-noise ratio environments, the visual textures of m-D spectrograms severely degrade, causing significant performance drops in conventional deep learning models. To address this issue, this paper proposes an end-to-end privileged information-based cross-modal fusion (PiBi) framework. First, a high-fidelity multi-scattering center dynamic model of the Beidou-2 satellite is constructed to generate a comprehensive m-D dataset comprising six typical micro-motion states. Subsequently, a dual-branch feature extraction network is designed to independently capture high-dimensional visual features and explicit physical statistical priors from the spectrograms. Finally, a compact gated bilinear fusion module is introduced to dynamically aggregate these heterogeneous features through secondorder interaction. Under the constructed Beidou-2 satellite micro-motion recognition task and simulated low signal-to-noise ratio scenarios, experimental results based on multiple modern visual backbones demonstrate that the proposed PiBi framework significantly enhances noise robustness and efficiently mitigates the inter-class confusion of weak-texture targets, maintaining high recognition accuracy even under signal-to-noise ratio conditions down to -18 dB.