[Paper Review] City-Wide Perceptions of Neighbourhood Quality using Street View Images

Emily Muller, Emily Gemmell+7 more|arXiv (Cornell University)|Nov 22, 2022

Urban Transport and Accessibility4 citations

TL;DR

This paper proposes a city-specific methodology for assessing urban residents' perceptions of neighbourhood walkability using street-level images and crowd-sourced ratings in London. By collecting 70,000 street-view images, training a deep learning model on user preferences, and mapping walkability scores across Greater London, the study enables high-resolution, perception-based urban planning with actionable insights for equity and regeneration.

ABSTRACT

The interactions of individuals with city neighbourhoods is determined, in part, by the perceived quality of urban environments. Perceived neighbourhood quality is a core component of urban vitality, influencing social cohesion, sense of community, safety, activity and mental health of residents. Large-scale assessment of perceptions of neighbourhood quality was pioneered by the Place Pulse projects. Researchers demonstrated the efficacy of crowd-sourcing perception ratings of image pairs across 56 cities and training a model to predict perceptions from street-view images. Variation across cities may limit Place Pulse's usefulness for assessing within-city perceptions. In this paper, we set forth a protocol for city-specific dataset collection for the perception: 'On which street would you prefer to walk?'. This paper describes our methodology, based in London, including collection of images and ratings, web development, model training and mapping. Assessment of within-city perceptions of neighbourhoods can identify inequities, inform planning priorities, and identify temporal dynamics. Code available: https://emilymuller1991.github.io/urban-perceptions/.

Motivation & Objective

• To develop a scalable, city-specific protocol for measuring residents' perceptions of urban neighbourhood quality using street-level imagery.
• To address the limitation of global perception datasets like Place Pulse by focusing on within-city variation in urban perception.
• To create a high-resolution, perception-based map of walkability across London using deep learning and crowd-sourced ratings.
• To support urban planning by identifying inequities, guiding regeneration efforts, and informing policy through user-centered metrics.
• To provide a reproducible framework for other cities to assess urban vitality through perceptual data derived from street views.

Proposed method

• Collected 70,000 street-view images from London at 20-meter intervals along roads to achieve ~70% street coverage.
• Designed a web application to host a pairwise image rating survey: 'On which street would you prefer to walk?'
• Used Amazon Mechanical Turk (AMT) to collect 1.2 million pairwise comparisons from 1,200+ participants for 7 perception dimensions.
• Trained a deep convolutional neural network (CNN) to predict walkability scores from image features, with class imbalance addressed via oversampling of extreme scores.
• Applied model-agnostic interpretability techniques to analyze features linked to high and low walkability perceptions.
• Generated spatially resolved maps of walkability scores across Greater London, with Output Areas color-coded by decile of average score.

Experimental results

Research questions

• RQ1 How can city-specific perceptions of urban walkability be systematically collected at scale using street-view imagery and crowd-sourcing?
• RQ2 To what extent do deep learning models trained on city-specific perception data generalize to predict human preferences for walkability in urban environments?
• RQ3 How do perceptual features such as greenery, street wall continuity, and urban design elements correlate with perceived walkability in London?
• RQ4 What is the spatial distribution of perceived walkability across London’s neighbourhoods, and where are the most and least walkable areas identified by residents?
• RQ5 How does the model’s performance vary across different demographic groups in the crowd-sourced annotator pool, and what are the implications for representativeness?

Key findings

• The study achieved ~70% street coverage in Greater London by sampling images at 20-meter intervals along road networks.
• The deep learning model demonstrated moderate performance in predicting walkability scores, with lower accuracy compared to other perception dimensions, likely due to subjective and context-dependent nature of walkability.
• Perception scores followed a normal distribution centered at μ=25, with oversampling of extreme values used to improve model learning of tail behaviors.
• Model-agnostic interpretability revealed that features such as greenery, building continuity, and pedestrian infrastructure were strongly associated with higher walkability perceptions.
• Spatial mapping revealed significant variation in walkability perceptions across London’s boroughs, with higher scores concentrated in central and suburban green corridors.
• The method enables high-resolution, perception-based urban mapping that can inform equitable urban planning and identify areas in need of regeneration or investment.