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

This study develops a set of algorithms to extract built environment features from Google aerial and street view images, reflecting the microcharacteristics of an urban location as well as the different functions of buildings. These features were used to train a Bayesian regularized artificial neural network (BRANN) model to predict near-road air quality based on measurements of ultrafine particles (UFPs) and black carbon (BC) in Toronto, Canada. The resulting models [adjusted R² of 75.87 and 79.10% for UFP and BC and root mean squared error (RMSE) of 21,800 part/cm³ and 1300 ng/m³ for UFP and BC] were compared with similar ANN models developed using the same predictors, but extracted from traditional geographic information system (GIS) databases [adjusted R² of 58.74 and 64.21% for UFP and BC and RMSE values of 23,000 part/cm³ and 1600 ng/m³ for UFP and BC]. The models based on feature extraction exhibited higher predictive power, thus highlighting the greater accuracy of the proposed methods compared to GIS layers that are solely based on aerial images. A comparison with other neural network approaches as well as with a traditional land-use regression model demonstrates the strength of the BRANN model for spatial interpolation of air quality.