Urban-scale facade material mapping from street view images using vision-language models for circular construction planning.

The lack of reliable building-level data remains an obstacle for advancing urban sustainability and circular economy practices. Here, we present URBAN-AI, a workflow that employs multimodal artificial intelligence to infer building material and facade characteristics from street-level imagery. We apply the workflow across six cities spanning high-income to low-income contexts (Zurich, San Francisco, Melbourne, Mumbai, Cape Town, and Rio de Janeiro), analysing and verifying outputs for 9,056 high confidence images, and achieving a mean composite module accuracy of 87.7% across all cities; to our knowledge, this represents the first cross-city evaluation of facade material classification using vision-language models. In all contexts, and especially in low-income and data-scarce settings, the workflow provides basic building information such as typology, material type, condition, and architectural style. The workflow also generates image-based indicators to inform downstream applications, including historical facade visual proxies, seismic retrofit visual proxies, energy retrofit visual proxies, urban morphology indicators, facade greening suitability visual proxies, and flood exposure visual proxies, to complement domain-specific models. We evaluate the approach in terms of accuracy, coverage, and cost efficiency, and contextualise results against conventional methods such as manual surveys. The study contributes two open resources: the Global Building Facade Dataset and the URBAN-AI workflow. These findings demonstrate the potential of AI-based workflows to support future efforts in circular construction, resource planning, and urban sustainability.