Predicting walking-to-work using street-level imagery and deep learning in seven Canadian cities

Area of Science:

• Urban planning
• Public health
• Computer vision

Background:

• Street-level imagery offers novel health-relevant urban environmental data.
• Previous studies on urban features and active commuting are limited in Canada.
• Google Street View (GSV) data provides extensive visual information for urban analysis.

Purpose of the Study:

• To extract urban environmental features from GSV images in seven Canadian cities.
• To assess the association between these features and walk-to-work rates.
• To compare the predictive performance of GSV-derived features against traditional walkability measures.

Main Methods:

• Utilized 1.15 million Google Street View images.
• Applied image segmentation and object detection computer vision techniques.
• Extracted data on persons, bicycles, buildings, sidewalks, open sky, and vegetation at the postal code level.

Main Results:

• GSV-derived urban features significantly predicted walk-to-work rates.
• Street-level image features outperformed traditional walkability metrics in predicting active commuting.
• Demonstrated the utility of computer vision for analyzing urban environments.

Conclusions:

• Street-level imagery and computer vision are powerful tools for urban health research.
• This methodology offers a more accurate approach to understanding active transportation patterns.
• Highlights the potential for machine learning in studying health behaviors and exposures.