AIRSAT-Bench: Optical-SAR Bimodal Remote Sensing Interpretation Benchmark Dataset

The AIRSAT-Bench (Optical-SAR Bimodal Remote Sensing Interpretation Benchmark Dataset) is a large-scale, pixel-level optical-SAR bimodal coregistered dataset designed for intelligent geospatial feature extraction and related tasks, aimed at advancing research in multi-source remote sensing collaborative interpretation.Leveraging the AIRSAT constellation developed by CAS Satellite (Zhongke Weixing), the dataset encompasses both optical and SAR satellite modalities, covering a typical agricultural area in Laiwu, Shandong Province. The SAR imagery is acquired from two on-orbit satellites, AIRSAT-05 and AIRSAT-08. Specifically, AIRSAT-05 operates in X-band with single, dual, and full polarization imaging capabilities, achieving a best resolution better than 1 meter; AIRSAT-08 operates in X-band with dual polarization, also achieving a best resolution better than 1 meter. The optical imagery is sourced from the on-orbit AIRSAT-07 satellite, with a best resolution better than 1 meter.The data annotation pipeline employs a four-tier quality control mechanism comprising "AI pre-labeling, manual refinement, cross-validation, and expert final review," encompassing seven typical land cover categories: maize, wheat, greenhouses, roads, buildings, rivers and lakes, and ponds. The AIRSAT-Bench dataset provides both original wide-swath annotated imagery and annotated image patches at three scales (256×256, 512×512, and 1024×1024 pixels); the original wide-swath annotated imagery is distributed in GeoTIFF format. The dataset comprises a total of 83,665 valid image-annotation pairs, covering optical single-modal, SAR single-modal, and optical-SAR bimodal coregistered configurations. Researchers may select either the original wide-swath annotated imagery or patch data at the available scales according to specific task requirements. The AIRSAT-Bench dataset offers high-quality benchmark support for diverse remote sensing interpretation tasks, including semantic segmentation, bimodal coregistration, and change detection.

AIRSAT-Bench（光学-合成孔径雷达双模态遥感解译基准数据集，Optical-SAR Bimodal Remote Sensing Interpretation Benchmark Dataset）是一款大规模像素级光学-合成孔径雷达（SAR）共配准双模态数据集，专为智能地理空间特征提取及相关任务设计，旨在推动多源遥感协同解译领域的研究进展。该数据集依托中国科学院卫星（中科卫星，CAS Satellite）研发的AIRSAT卫星星座构建，涵盖光学与合成孔径雷达（SAR）两种卫星模态，覆盖山东省莱芜市一处典型农业区域。合成孔径雷达（SAR）影像取自两颗在轨卫星AIRSAT-05与AIRSAT-08。其中，AIRSAT-05工作于X波段，具备单极化、双极化及全极化成像能力，最优分辨率优于1米；AIRSAT-08同样工作于X波段，支持双极化成像，最优分辨率亦优于1米。光学影像则来自在轨卫星AIRSAT-07，其最优分辨率优于1米。该数据集的标注流程采用“AI预标注、人工精修、交叉验证、专家终审”四级质量控制机制，覆盖玉米、小麦、温室大棚、道路、建筑、河湖与池塘共7类典型土地覆盖类别。AIRSAT-Bench数据集同时提供原始宽幅标注影像与三种尺度（256×256、512×512及1024×1024像素）的标注图像切块；原始宽幅标注影像采用GeoTIFF格式分发。该数据集总计包含83665组有效图像-标注对，涵盖光学单模态、合成孔径雷达（SAR）单模态以及光学-合成孔径雷达（SAR）双模态共配准三种数据配置形式。研究人员可根据具体任务需求，选择原始宽幅标注影像或对应尺度的图像切块数据。AIRSAT-Bench数据集可为语义分割、双模态共配准、变化检测等多种遥感解译任务提供高质量基准支撑。