Investigating Links Between Urban Residential Streetscapes and Physical Activity Using Deep Learning of Google Street View Imagery Applied to the Washington State Twin Registry.

Abstract

The evidence linking urban greenspace to individual's physical activity (PA) levels is mixed. This study examines relationships between street-level and satellite-derived greenspace measures with PA outcomes. Our sample included 7855 adult twins enrolled in the Washington State Twin Registry from 2009 to 2020 living in urban areas; 14,095 total survey observations were analyzed. We applied a deep learning segmentation algorithm to Google Street View images sampled from 100 m around residential addresses to quantify street-level greenspace. Bouts and duration of PA, including moderate to vigorous PA and neighborhood walking were self-reported. We applied mixed-effects linear regression models to determine relationships between greenspace measures and PA outcomes, overall and stratified by residential population density. Adjusted models included age, body mass index, sex, race, education, income, neighborhood deprivation, urban sprawl, and seasonality. A series of sequential models was constructed to test associations between various greenspace exposures and PA outcomes. Overall, we found no consistent associations between greenspace exposures and PA outcomes. We found that the summer normalized difference vegetation index was associated with an increase in moderate to vigorous PA in low population density areas, but this was not significant when controlling for seasonality. Both Google Street View and normalized difference vegetation index were associated with lower total walking for those residing in areas with high population density only. Findings highlight the importance of seasonality and the need to address where PA is actually done.