The intangible value of street-level urban design using a machine learning approach: urban beauty, active frontage and street greener
Can the visual qualities of neighbourhoods predict property values? Recent advances in machine learning offer new ways to quantify urban design at scale, but often produce results that are difficult to interpret or act upon.
Law, S., Seresinhe, C. I. In preparation
We extract three established urban design features from street-level imagery and test their ability to predict house prices. Unlike previous approaches that rely on opaque deep learning features, we focus on interpretable design elements that urban planners already understand. We also compute neighbourhood-level variants of these features, recognising that property values reflect not just individual streets but entire areas.
We find that these theory-driven urban design features not only improve house price predictions compared to models without visual data, but perform competitively against state-of-the-art deep learning approaches (VGG16 and ViT). Crucially, because our features correspond to real design elements—like building setbacks, street enclosure, and visual complexity—they can directly inform housing and urban design policy. This work demonstrates that we don't always need to sacrifice interpretability for accuracy when applying AI to urban challenges.