BuildingMultiView:powering multi-scale building characterization with large language models and Multi-perspective imagery

作者：Zongrong Li, Yunlei Su, Filip Biljecki, Wufan Zhao · 发表于：International Journal of Applied Earth Observation and Geoinformation · 年份：2026 · DOI：10.1016/j.jag.2025.105034 · 被引用次数：2 · 研究领域：Human Mobility and Location-Based Analysis、Advanced Neural Network Applications、Remote Sensing and LiDAR Applications

Buildings play a crucial role in shaping urban environments, influencing their physical, functional, and aesthetic characteristics. However, urban analytics is frequently limited by datasets lacking essential semantic details as well as fragmentation across diverse and incompatible data sources. To address these challenges, we conducted a comprehensive meta -analysis of 6,285 publications (2019–2024). From this review, we identified 11 key visually discernible building characteristics grouped into three branches: satellite house, satellite neighborhood, and street-view. Based on this structured characteristic system, we introduce BuildingMultiView, an innovative framework leveraging fine-tuned Large Language Models (LLMs) to systematically extract semantically detailed building characteristics from integrated satellite and street-view imagery. Using structured image–prompt–label triplets, the model efficiently annotates characteristics at multiple spatial scales. These characteristics include swimming pools, roof types, building density, wall–window ratio, and property types. Together, they provide a comprehensive and multi-perspective building database. Experiments conducted across five cities in the USA with diverse architecture and urban form, San Francisco, San Diego, Salt Lake City, Austin, and New York City, demonstrate significant performance improvements, with an F1 score of 79.77% compared to the untuned base version of ChatGPT’s 45.66%. These results reveal diverse ...