Deep learning meets urban design: Assessing streetscape aesthetic and design quality through AI and cluster analysis

Abstract

Urban streetscapes possess unique aesthetic qualities, yet the link between street design and aesthetic quality, particularly across different streetscape types, remains underexplored due to evaluation limitations. This study bridges the gap by integrating deep learning and cluster analysis. We constructed a dataset through manual aesthetic ratings of 500 sampled street images by 150 participants, then fine-tuned a ResNet-50 to automatically predict the aesthetic of 16,455 streetscape images from Wuhan. Street design quality was assessed using visual landscape elements derived from semantic segmentation, while the HDBSCAN algorithm was employed to cluster streetscape types. Multiple linear regression and grouped regression analysis were employed to examine design-aesthetic relationships across six streetscape types. Key findings include: (1) the model's high accuracy in assessing aesthetics, with vegetation and water as key factors; (2) streetscapes were categorized into six clusters, with highways and bridges rated as having the highest aesthetic quality, and overpasses rated the lowest; (3) greenness, openness, and walkability strongly correlated with aesthetics, especially in residential streets; and (4) The impact of street design quality on aesthetic perception varies significantly across different types of streets, suggesting that street function may influence aesthetic standards. This study introduces a novel framework for evaluating urban streetscape aesthetics and design, providing actionable insights for creating sustainable, aesthetically pleasing environments.