As the strategic importance of space continues to grow, space target recognition technology is critical for advancing space surveillance systems, optimizing the use of space resources, and ensuring space security. In response to challenges in analyzing radar cross-section (RCS) sequences, we propose a deep learning-based method designed to enhance the robustness and accuracy of RCS sequence analysis for space target recognition. We introduce a novel period estimation method based on combination functions and analysis of variance (ANOVA), which effectively suppresses noise and captures periodic characteristics with greater accuracy. Building on this, we propose a Transformer-based approach for size estimation from RCS sequences, leveraging Transformers’ advanced sequence modeling to reduce common errors in traditional methods, further improving space target characterization. We then integrate period and size features into a unified feature set and introduce a cross-attention-based multi-feature interaction module to fuse physical and statistical features, learning dependencies between them to enhance target recognition accuracy. Experimental results demonstrate that our approach significantly improves both performance and stability of space target recognition, providing a solid foundation for further advancements in space surveillance technology.