中国国家级多属性建筑数据集CMAB

快速获取三维（3D）建筑数据（包括屋顶、高度、朝向等几何属性，以及功能、质量、年代等指示性属性）对精准城市分析、模拟和政策更新至关重要。针对现有建筑数据集多属性覆盖不全的缺陷，本研究通过人工智能技术构建了首个国家级多属性建筑数据集（CMAB），覆盖3,667个自然城市、3,100万栋建筑和236亿平方米屋顶（基于OCRNet提取的F1值达89.93%），建筑存量总体积达3,630亿立方米。通过城市行政分类训练自举聚合XGBoost模型，融合形态、区位与功能特征，结合百亿级遥感影像与6,000万张街景图像（SVIs）等多源数据，利用机器学习与大型多模态模型为每栋建筑生成屋顶、高度、结构、功能、风格、年代和质量属性。精度验证通过模型基准测试、同类产品对比和人工街景验证完成，多数指标超80%。本数据集对全球可持续发展目标（SDGs）与城市规划具有关键价值。
涵盖中国3,667个自然城市、总屋顶面积236亿平方米的建筑数据集，包含建筑屋顶、高度、结构、功能、年代、风格、色彩与质量属性，以及数据计算所用的代码文件。使用的深度学习模型包括OCRNet、XGBoost、微调CLIP和Yolo-v8。

Rapid access to three-dimensional (3D) building data—including geometric attributes such as roof shape, height, and orientation, as well as indicative attributes like function, quality, and construction era— is critical for precise urban analysis, simulation, and policy updates. Aiming to address the gap of incomplete attribute coverage in existing building datasets, this study constructs the first national-scale multi-attribute building dataset (CMAB) using artificial intelligence technologies. It covers 3,667 natural cities in China, 31 million buildings, and 23.6 billion square meters of roof area (with an F1 score of 89.93% achieved via OCRNet extraction), with a total building volume of 363 billion cubic meters. By training a bootstrap aggregated XGBoost model based on urban administrative classification, integrating morphological, locational, and functional features, and combining multi-source data including billions of remote sensing images and 60 million street view images (SVIs), this work leverages machine learning and large multimodal models to generate attributes including roof features, height, structure, function, style, construction era, and quality for each individual building. Accuracy validation was conducted via model benchmark tests, comparisons with existing similar products, and manual street view verification, with most metrics exceeding 80%. This dataset is of critical value for the United Nations Sustainable Development Goals (SDGs) and urban planning.

This building dataset covers 3,667 natural cities across China, with a total roof area of 23.6 billion square meters. It includes attributes such as building roof, height, structure, function, construction era, style, color and quality, as well as the code files used for data computation. The deep learning models employed include OCRNet, XGBoost, fine-tuned CLIP, and YOLOv8.

搜集汇总

背景与挑战

背景概述

中国国家级多属性建筑数据集CMAB是中国首个国家级多属性建筑数据集，覆盖3667个自然城市、3100万栋建筑物，总屋顶面积236亿平方米。该数据集通过人工智能技术生成，包含建筑物的屋顶、高度、结构、功能、年代、风格和质量等多维属性，大部分指标准确率超过80%，对可持续发展和城市规划具有重要价值。