Assessing walkability with deep CNNs by integrating objective and subjective urban qualities: The case of ‘Cittaslow’ neighborhoods

Walkability, shaped by features of built environment, contributes to healthier communities and more sustainable urban mobility. Well-designed streetscapes encourage pedestrian activity, making individuals’ perceptions of safety, accessibility, and aesthetics central to understanding walkable environments. Existing studies typically evaluate walkability through either objective spatial data or perceptual assessments based on human experiences. However, comprehensive approaches that combines both perspectives remain limited. This study presents an AI-driven framework for evaluating urban walkability by integrating objective built environment features with subjective perceptions in the context of Cittaslow-certified neighborhoods. The research employs the Segment Anything Model 2(SAM2) for high-resolution and class-agnostic segmentation of street-level imagery. It is created: “Urban Walkability Dataset” (UWD) which contains 5,440 labeled images by experts with a question set generated based on the key parameters affecting walkability. A neural network pipeline is designed to understand the underlying process. By bridging perceptual insights and objective metrics, this research contributes a replicable methodology for walkability assessment that supports human-centered urban design strategies, particularly in slow-city contexts prioritizing sustainability and quality of life.