Graph-based machine learning for high-resolution assessment of pedestrian-weighted exposure to air pollution

Abstract

Pedestrians are particularly vulnerable to air pollution due to their proximity to pollutant sources and elevated respiratory rates during physical activity, amplifying cumulative health risks. However, existing studies focus on concentration- or residence-based exposure assessment, overlooking the dynamic interaction between pollution patterns and pedestrian activity. This study therefore introduces a novel methodological framework to assess pedestrian-specific exposure to PM2.5 in diverse urban environments. Applied to New York City, the framework leverages graph-based machine learning to predict street-level PM2.5 concentrations from vehicle-sensed pollution data, while estimating high-resolution pedestrian volume derived from street view imagery and ground-truth count data. The results reveal significant divergences between traditional exposure assessments and pedestrian-specific exposure patterns, uncovering previously overlooked high-risk zones. High-exposure hotspots are not limited to areas with elevated pollution levels but also include locations where moderate pollution coincides with high pedestrian activity. This study also explores the spatial relationship between exposure patterns and urban vegetation coverage, providing actionable insights for targeted interventions. By bridging the gap between pollution dynamics and pedestrian activity, this research provides urban planners and policymakers with new insights for developing pedestrian-centered air quality management strategies, contributing to healthier and more sustainable urban environments.