FAIR-CSAR复图像目标数据集

数据主编：张腊梅（哈尔滨工业大学） 数据集简介：FSAR-Cap 数据集旨在为SAR图像语义理解与跨模态建模构建一套具备精细语义描述能力的图像–文本对照语料，推动 SAR图像自动解译、图像描述生成（Captioning）以及遥感多模态模型的发展。该数据集基于FAIR-CSAR的目标检测数据集构建，共包含14,480张SAR图像和72,400条配套描述文本。FSAR-Cap采用两阶段标注方式：先利用检测结果和空间位置信息，通过多种模板自动生成基础描述；随后再结合人工校验与语言大模型润色。最终，每幅图像都会生成5条风格不同、信息互补的描述语句，文本内容涵盖目标类型、数量、位置关系、外形特征等。 作为目前首个面向SAR图像、具备细粒度描述层级的大规模语义标注数据集，FSAR-Cap不仅提升了SAR图像语义表达质量，也为SAR领域的图像captioning、遥感视觉语言模型训练、多模态推理、SAR–自然语言对齐研究提供了统一且高质量的数据基准，为SAR自动化解译与智能语义理解技术体系的进一步发展奠定了基础。 详细使用说明请参考：FSAR-Cap：大规模细粒度SAR图像描述使用说明.pdf 本数据集参考文献与引用格式： [1] 张金琪, 庄迪, 张腊梅等. FSAR-Cap：大规模细粒度SAR图像描述数据集[OL]. 雷达学报, 2026. https://radars.ac.cn/web/data/getData?dataType=FAIR_CSAR. ZHANG Jinqi, ZHUANG Di, ZHANG Lamei, et al. FSAR-Cap：Large-scale fine-grained SAR image captioning dataset [OL]. Journal of Radars, 2026. https://radars.ac.cn/web/data/getData?dataType=FAIR_CSAR_en&pageType=en. [2] 张金琪, 庄迪, 张腊梅等. DGS-CapNet：基于空间-频率感知的SAR图像描述模型[J]. 雷达学报(中英文), 待出版. doi: 10.12000/JR25250. ZHANG Jinqi, ZHUANG Di, ZHANG Lamei, et al. DGS-CapNet: A Spatial-Frequency-Aware Model for SAR Image Captioning[J]. Journal of Radars, in press. doi: 10.12000/JR25250. 发布日期：2026年1月7日

Dataset Editor: Lamei Zhang (Harbin Institute of Technology)
Dataset Introduction: The FSAR-Cap dataset aims to construct a high-quality image-text aligned corpus with fine-grained semantic description capabilities for SAR image semantic understanding and cross-modal modeling, to promote the development of SAR image automatic interpretation, image captioning, and remote sensing multimodal models. This dataset is built based on the FAIR-CSAR object detection dataset, containing a total of 14,480 SAR images and 72,400 accompanying descriptive texts. FSAR-Cap adopts a two-stage annotation pipeline: firstly, basic descriptions are automatically generated via multiple templates using detection results and spatial location information; subsequently, manual verification and polishing by large language models (LLMs) are conducted. Finally, 5 descriptive statements with distinct styles and mutually complementary information are generated for each image, with the text covering target types, quantities, positional relationships, morphological features, and other relevant details.
As the first large-scale semantic annotation dataset for SAR images with fine-grained description hierarchy to date, FSAR-Cap not only improves the quality of semantic expression for SAR images, but also provides a unified and high-quality data benchmark for SAR-domain image captioning, remote sensing vision-language model training, multimodal reasoning, and SAR-natural language alignment research, laying a solid foundation for the further advancement of SAR automatic interpretation and intelligent semantic understanding technology systems.
For detailed usage instructions, please refer to: FSAR-Cap: User Guide for Large-scale Fine-grained SAR Image Captioning.pdf
References and Citation Formats:
[1] ZHANG Jinqi, ZHUANG Di, ZHANG Lamei, et al. FSAR-Cap: Large-scale Fine-grained SAR Image Captioning Dataset [OL]. Journal of Radars, 2026. https://radars.ac.cn/web/data/getData?dataType=FAIR_CSAR_en&pageType=en
[2] ZHANG Jinqi, ZHUANG Di, ZHANG Lamei, et al. DGS-CapNet: A Spatial-Frequency-Aware Model for SAR Image Captioning[J]. Journal of Radars, in press. doi: 10.12000/JR25250
Release Date: January 7, 2026