Using Street View Imagery to Predict Street-Level Particulate Air Pollution

Area of Science:

• Environmental Science
• Computer Science
• Public Health

Background:

• Traditional land-use regression (LUR) models for air pollution estimation use fixed-site data and limited-resolution GIS variables.
• There's a need for higher spatial resolution air pollution models to identify localized pollution sources and impacts.

Purpose of the Study:

• To develop and evaluate an approach using Google Street View (GSV) imagery and deep learning to predict street-level black carbon (BC) and particle number (PN) concentrations.
• To compare the performance of GSV-based LUR models with traditional LUR models and assess the impact of feature selection and buffer sizes.

Main Methods:

• Collected mobile monitoring data for BC and PN concentrations.
• Extracted features from approximately 52,500 GSV images using a deep learning model.
• Developed empirical LUR models with varying buffer sizes (50-2000 m) and tested feature selection methods.

Main Results:

• GSV-based models achieved comparable performance to traditional LUR models, with adjusted R-squared values of 0.57-0.64 for BC and 0.65-0.73 for PN.
• Models using GSV images within 250 m explained ~50% of air pollution variability, highlighting the influence of the immediate built environment.
• The approach identified additional potential pollution hotspots.

Conclusions:

• Google Street View imagery, processed with computer vision, is a valuable data source for developing high-resolution LUR models.
• This method provides consistent predictor variables across administrative boundaries, overcoming limitations of traditional approaches.
• The findings support the use of GSV data for more accurate street-level air pollution assessment.