Exploring nonlinear and interactive associations between built environment features and subjective streetscape perceptions

This study explores nonlinear and interactive relationships between built environment factors and six subjective streetscape perceptions (i.e., beauty, wealth, safety, liveliness, boredom, and depression) in Jeonju, South Korea. Employing a two-stage methodological framework, we first estimate the perceptual scores using a Convolutional Neural Network trained on Place Pulse data and Street View images, and then utilize a Gradient Boosting Regressor with interpretable machine learning methods to examine the complex relationships. Key findings reveal several patterns. First, although perceptions of beauty and safety are generally positively associated with housing prices, the association with safety becomes marginally negative in neighborhoods with exceptionally high housing values. Furthermore, interaction effects highlight that areas characterized by both a higher proportion of older buildings and elevated housing prices are more likely to be perceived as beautiful. Conversely, neighborhoods with simultaneously low population and employment densities are associated with higher boredom scores. This study contributes to (1) demonstrating the utility of AI-based approaches in the field of urban planning and (2) deepening theoretical and empirical insights into how urban form shapes subjective experience from the urban shrinkage perspective.