遥感影像建筑物信息提取对城市信息管理与防灾减灾等领域具有至关重要的作用。本文建立了一个高分辨率多光谱遥感影像细粒度建筑物特征集MFBFS。MFBFS采用国产高分二号多光谱遥感影像作为数据源，选择覆盖了总共3668平方千米的中国各个灾害带的21个区县建筑物集中区域为研究区，从光谱、纹理、边缘、指数四个角度，生成了17种特征分量。MFBFS中共包含超过26万个建筑物实例，大规模的数据保证了较高的类内差异，包括尺寸、形状、颜色、角度、背景等差异，为后续泛化性模型的建立提供了数据支撑。此外，MFBFS特有地将建筑物按照结构类型分为钢及钢筋混凝土结构、砌体结构以及砖石和其他结构三种。不同结构类型的建筑物抵御灾害的能力以及可使用时间区别显著，细粒度的设计使得遥感提取建筑物任务将发挥更大的作用，尤其是灾害领域的灾前损失预测和灾后损失评估。为保证地面真实值的高度准确性，我们进行了严格的质量流程控制和实地考察验证工作，最终得到191GB高质量特征和标签数据。初步的深度学习实验表明了MFBFS的有效性。该特征集(下载方式：https://github.com/WangZhenqing-RS/MFBFS） 可为建筑物结构细粒度提取研究提供良好的数据支持，也可促进国产高分遥感数据应用发展。

Building information extraction from remote sensing images plays a vital role in urban information management and disaster prevention and mitigation. This paper establishes a fine-grained building feature set MFBFS for high-resolution multispectral remote sensing images. MFBFS uses the domestically produced Gaofen-2 multispectral remote sensing images as the data source, and selects 21 districts and counties with concentrated buildings in various disaster zones in China covering a total of 3,668 square kilometers as the study area. From four perspectives: spectrum, texture, edge, and index, 17 feature components are generated. MFBFS contains a total of more than 260,000 building instances. The large-scale data ensures high intra-class differences, including differences in size, shape, color, angle, background, etc., providing data support for the establishment of subsequent generalization models. In addition, MFBFS uniquely divides buildings into three types according to structural types: steel and reinforced concrete structure, masonry structure, and block stone structure. There are significant differences in the ability of buildings of different structural types to withstand disasters and their usable time. The fine-grained design will make the task of extracting buildings through remote sensing play a greater role, especially in the pre-disaster loss prediction and post-disaster loss assessment in the disaster field. To ensure the high accuracy of ground truth values, we conducted strict quality process control and on-site inspection verification work, and finally obtained 191GB of high-quality feature and label data. This data set (download method: https://github.com/WangZhenqing-RS/MFBFS) can provide good data support for fine-grained extraction research on building structures, and can also promote the application development of domestic high-resolution remote sensing data.