Multi-Modal Large Models Driven SAR Image Captioning: A Benchmark Dataset and Baselines

Despite the remarkable progress achieved by synthetic aperture radar (SAR) imagery in tasks such as scene classification and object detection in recent years, automatically generating accurate and structurally coherent natural language descriptions remains a significant challenge due to the inherent structural complexity and semantic ambiguity of SAR data. In this paper, we propose the first large-scale benchmark dataset tailored for the SAR image captioning task. Our central insight is that a hierarchical semantic modeling strategy can effectively bridge the gap between the unique visual characteristics of SAR imagery and natural language expression. To this end, we design an annotation protocol that spans three semantic layers—object, relation, and scene—and develop a semi-automated annotation framework assisted by the large model GPT-4V, significantly reducing human annotation cost while ensuring annotation quality. Based on this strategy and framework, we construct the ATRNet-SARCap 1.0 dataset, which focuses on aircraft targets and covers 5,251 images and 47,259 structured caption sentences. We further conduct systematic experiments using various mainstream encoder-decoder frameworks for visionlanguage modeling. The results demonstrate that our dataset substantially enhances semantic completeness and expression accuracy in the SAR image captioning task, providing a highquality data foundation and research paradigm for advancing language understanding of remote sensing imagery.