Work

Spatial Predictions of Active Travel Based on Google Street View Imagery and Place of Interest Data

Year: 2019

Type: article

Abstract: OPS 58: Predictors and impact of active travel, Room 210, Floor 2, August 28, 2019, 10:30 AM - 12:00 PM Background/Aim: Increasing active travel is a commonly cited strategy for improving health in urban areas. Few studies assess trends of walking and cycling across multiple cities. Most studies use jurisdiction-specific land use data at the neighborhood-level. We model bicycle and pedestrian traffic at 6,342 count locations across 20 US cities using machine learning approaches and integrating new Google-derived, street-level variables. Methods: We build on previous models of active travel by developing a new set of Google-derived variables. These variables include (1) information on destinations (e.g., restaurants, banks, schools, hospitals) from Google Place of Interest (POI) data and (2) object identification using deep learning models from Google Street View (GSV) imagery (e.g., trees, street amenities, vehicles, buildings, water bodies). We merged these new datasets with traditional Census variables (tabulated using various buffer distances to represent neighborhood characteristics). We compared use of stepwise linear regression to machine learning to assess which variables contributed most to improved model performance (e.g., Census-based vs. Google-derived variables and neighborhood-level vs. street-level variables). Results: We found that machine learning (as compared to stepwise linear regression) improved model prediction when using Census-based, neighborhood-level variables (10-fold CV R2: 0.86-0.89). Models that used only Google-derived variables had similar model performance (10-fold CV R2: 0.80-0.81) to the traditional Census variables. GSV variables were more important in the bicycle models as compared to the pedestrian models where POI variables were more frequently selected. This finding suggests that street-level interventions may be more effective on bicycles while neighborhood-level interventions may be needed to promote walking. Models that included both traditional and Google-derived variables performed similarly to models that used each set of variables individually. Conclusions: Our results suggest that Google-derived, street-level variables may be a useful alternative to traditional neighborhood-level variables towards improved prediction of bicycle and pedestrian traffic volumes.

Source: Environmental Epidemiology

DOI: 10.1097/01.ee9.0000607440.14860.a9

Authors: Hankey S, Zhang W, Le H, Hystad P, James P

Institutions: Virginia Tech, Oregon State University, Harvard University, Harvard Pilgrim Health Care

Language: English

Topic: Urban Transport and Accessibility

Subfield: Transportation

Field: Social Sciences

Domain: Social Sciences

SDGs: Sustainable cities and communities, Good health and well-being, Industry, innovation and infrastructure

Open Access status: gold

Estimated APC paid: $1,200