Predicting Traffic-Related Air Pollution Using Feature Extraction from Built Environment Images.

Area of Science:

• Environmental Science
• Urban Planning
• Data Science

Background:

• Urban built environments significantly influence local air quality.
• Traditional methods using Geographic Information System (GIS) data have limitations in capturing fine-grained environmental details.

Purpose of the Study:

• To develop and validate algorithms for extracting detailed built environment features from street-level imagery.
• To build and assess a Bayesian regularized artificial neural network (BRANN) model for predicting near-road air quality (ultrafine particles and black carbon).

Main Methods:

• Utilized Google aerial and street view images to extract micro-level urban characteristics and building functions.
• Trained a BRANN model using extracted features to predict ultrafine particle (UFP) and black carbon (BC) concentrations.
• Compared the BRANN model's performance against traditional GIS-based models and other neural network approaches.

Main Results:

• The BRANN model utilizing extracted features achieved higher predictive accuracy (adjusted R² of 75.87% for UFP, 79.10% for BC) compared to GIS-based models (adjusted R² of 58.74% for UFP, 64.21% for BC).
• The proposed feature extraction method demonstrated superior accuracy over traditional GIS layers.
• The BRANN model showed significant strength in the spatial interpolation of air quality data.

Conclusions:

• Algorithms extracting built environment features from street view images enhance the accuracy of air quality prediction models.
• The BRANN model, combined with detailed feature extraction, offers a powerful tool for understanding and predicting near-road air quality.
• This approach provides a more accurate alternative to traditional GIS-based spatial interpolation for air quality assessment.