Prototype-Driven Semantic Tree with Bimodal Embedding Space for Zero-Shot SAR Ship Recognition

Zero-shot SAR ship recognition aims to identify unseen ship categories without annotated SAR samples, offering a promising solution for open-category maritime remote sensing. Although synthetic aperture radar (SAR) provides all-weather and day-and-night observation capability, practical maritime surveillance often encounters newly emerging or rarely observed ship types with limited labeled data. Existing zero-shot recognition methods mainly rely on direct visual–semantic mapping, while overlooking the scattering-driven structural characteristics of SAR imagery and the semantic relationships among ship categories. This leads to two key challenges: weak alignment between SAR visual features and textual attributes, and semantic isolation of unseen categories in the embedding space. To address these issues, we propose a prototype-driven semantic tree framework with a bimodal embedding space for zero-shot SAR ship recognition. First, a Bimodal Embedding Space Construction (BESC) module is designed to learn structure-aware visual embeddings from SAR images and dependency-aware semantic embeddings from textual attributes, and align them within a unified embedding space. Second, a Prototype-Driven Semantic Tree (PDST) module organizes class prototypes into a hierarchical structure and refines them through tree-guided propagation, enabling structured knowledge transfer among related ship types. During inference, an unseen SAR image is classified by matching its visual embedding with the refined semantic prototypes. Experiments on the FUSAR and SRSDD datasets show that BESC-PDST improves the harmonic mean from 0.424 to 0.448 and from 0.509 to 0.541, respectively, outperforming representative zero-shot recognition methods. These results demonstrate the effectiveness of the proposed framework for knowledge-driven open-category SAR ship understanding.