Lightweight Spatial–Frequency Constraint Propagation Framework for Satellite Detection in Space Surveillance

Satellite object detection in space surveillance is challenged by sparse and weak targets in large-scale, structured backgrounds (e.g., star fields, clouds, and streak noise). Such interference is not random but exhibits spatial correlation and frequency regularity, causing target responses to be overwhelmed and difficult to separate within a single representation space. To address this issue, we propose a lightweight framework, termed DRSS-Net, based on the key observation that target–background separability can be enhanced across complementary representation coordinate systems. Specifically, spatial modeling captures local structural consistency, while frequency-domain processing characterizes global energy distribution and structured patterns. By alternating between these domains, the proposed method enables constraint propagation, where predictable background patterns are suppressed, and structurally inconsistent target responses are emphasized. In the spatial domain, a mutual conditioning mechanism with asymmetric channel allocation enhances the consistency between localization and semantic responses. In the frequency domain, a coupled refinement module models the interaction between energy distribution and structural configuration to distinguish structured background from anomalous targets. In addition, a scale selection strategy retains stable intermediate representations for efficient detection. Experiments on two independent space target datasets demonstrate that DRSS-Net consistently achieves superior detection performance with a compact model size under diverse observation conditions, including variations in target appearance, illumination, and structured background interference.