City Image: a dynamic perspective using machine learning and natural language processing

Abstract

The city image is a collective mental image of the city elements that can be perceived and interpreted by the public. The broader one’s understanding of the city image, the better urban design we will have. Over half a century ago, Kevin Lynch innovatively introduced this idea and summarized five physical elements – node, path, edge, landmark, and district, guiding the practice of urban design. Many subsequent studies confirmed its stability and further expanded the physical elements from a static viewpoint. However, cities are complex adaptive systems that involve both physical and subjective (affinity and reactions) aspects of temporal change.

This thesis focuses on a dynamic perspective of not only physical but also subjective city images during both day and night, and in different timeframes. Taking advantage of machine learning, this thesis measures how the public values the city based on hundreds of thousands of geo-tagged photos and their textual descriptions. The thesis demonstrates the possibility of large-scale studies on the city image. To identify its subjective associations, natural language processing is applied to extract frequently used words and evaluate sentiment analysis, reflecting the public’s affinity and reactions, positive and negative. Results are presented in the form of data visualization maps and charts. Case studies examined two major US cities and their representative elements – Boston (Fort Point Channel and Boston Common) and New York City (World Trade Center, High Line, and Brooklyn Bridge Park). The main conclusion is that there does exist the subjective city image based on dynamic analysis of Lynch’s physical elements which plays a key role in an in-depth understanding of city image. Based on the state-of-the-art technologies and perspective, the thesis sheds light on a comprehensive understanding of city image, formulating a new criterion as a potential guide for urban planning.