Developing Nationwide Estimates of Built Environment Quality Characteristics Using Street-View Imagery and Computer...

Area of Science:

• Environmental Health
• Urban Planning
• Computer Vision

Background:

• Built environment quality significantly impacts health behaviors, yet objective measures are lacking.
• Existing urban composition measures do not fully capture the nuances of environmental quality relevant to health.

Purpose of the Study:

• To develop and validate computer vision models for assessing multiple dimensions of built environment quality using street-view imagery.
• To generate nationwide, temporally resolved estimates of urban environmental quality across U.S. cities.

Main Methods:

• Collected 72,516 surveys via Amazon Mechanical Turk for training and validation.
• Utilized deep learning models to predict perceived beauty, relaxation potential, nature quality, safety for walking, and safety from crime.
• Adjusted for socio-demographic and seasonal biases in model predictions.

Main Results:

• Models achieved cross-validation accuracy from 59% (safety from crime) to 73% (nature quality), outperforming random chance (50%).
• Demographic adjustments partially reduced biases for specific ethnic groups.
• Seasonal bias correction improved model reliability for environmental quality metrics.

Conclusions:

• Nationwide street-level quality estimates (beauty, relaxation, nature, walking safety) are now available.
• These data can inform epidemiological studies, urban planning, and public health interventions.
• Safety from crime metric requires further refinement due to lower accuracy.