Global and local feature extraction of urban historical spatial perception using large language models: A case study of Harbin Central Street District

Spatial perception—the process by which individuals interpret and interact with their spatial environment—is a cornerstone of urban studies, forming the foundation for effective urban planning and policy-making. This study introduces innovative, digital methodologies for spatial perception analysis by integrating advanced technologies, including Large Language Models (LLMs) and BERTopic modeling. Using Harbin Central Street District, a historically significant urban area celebrated for its architectural heritage and scenic landscapes, as a case study, this research develops a replicable and scalable workflow for analyzing spatial perception. The study identifies global spatial perception patterns, such as pedestrian dynamics, green spaces, and cultural landmarks, alongside local themes unique to Central Street and Stalin Park. Key contributions include leveraging LLMs to generate synthetic text data from urban imagery, applying high-dimensional semantic vector models for topic analysis, and utilizing dynamic modeling techniques to extract both global and local spatial features. The findings reveal how historical architecture, commercial vibrancy, and natural environments interact to shape urban experiences, aligning closely with established spatial metrics. This research provides actionable insights for urban planning and policy, emphasizing heritage preservation, pedestrian-oriented design, and adaptive strategies. By bridging AI-driven technologies with urban research, this study offers a scalable, data-driven framework for analyzing spatial perceptions, advancing the digital transformation of cities, and promoting sustainable, inclusive urban development in the AI era.