CMAB: A Multi-Attribute Building Dataset of China

Area of Science:

• Geoinformatics
• Urban Analytics
• Artificial Intelligence

Background:

• Accurate 3D building data is crucial for urban analysis, simulations, and policy, but current datasets lack comprehensive multi-attribute coverage.
• Existing building datasets often have incomplete geometric and indicative attributes, limiting their utility for detailed urban studies.

Purpose of the Study:

• To present the first national-scale Multi-Attribute Building dataset (CMAB) with AI-driven, comprehensive building information.
• To provide a valuable resource for accurate urban analysis, simulations, policy updates, and global Sustainable Development Goals (SDGs).

Main Methods:

• Developed a national-scale dataset (CMAB) covering 3,667 cities and 31 million buildings using AI and machine learning.
• Employed OCRNet for attribute extraction (F1-Score 89.93%) and bootstrap aggregated XGBoost models incorporating morphology, location, and function.
• Utilized multi-source data, including remote sensing and street view images (SVIs), to generate rooftop, height, structure, function, style, age, and quality attributes.

Main Results:

• Generated a comprehensive national building dataset (CMAB) with extensive geometric and indicative attributes.
• Achieved high accuracy in attribute extraction, with an F1-Score of 89.93% for OCRNet and generally above 80% validation accuracy.
• Quantified building stock at 363 billion m³, providing unprecedented detail for urban research.

Conclusions:

• The CMAB dataset is a significant advancement for urban planning and analysis, addressing limitations of previous datasets.
• The AI-driven methodology demonstrates a scalable and accurate approach to generating rich building information.
• This dataset is vital for supporting global SDGs and informed urban development strategies.