Quantifying the effects of urban fabric and vegetation combination pattern to mitigate particle pollution in near-road areas using machine learning

Abstract

There are many intertwined factors such as urban fabric, vegetation combination, and meteorological conditions that influence the spatial distribution of fine particles (PM2.5, aerodynamic diameter less than 2.5 μm) in near-road urban areas, and it is difficult yet important to disentangle them. Identifying the relative importance of those factors could facilitate designing urban fabric and roadside vegetation combination patterns to mitigate the negative impacts of particles. This study first proposes an effective and reliable integrated method of the ENVI-met microclimate model and interpretable machine learning to quantify the multifactorial effects. Data from ENVI-met models with different built environments and meteorological conditions are used to train and test the machine learning (ML) models. All ML models show high accuracy in predicting PM2.5 concentrations with R2 high than 0.87. Results indicate that height level and study area are the two most important features, explaining 22.8% and 20.4% of the variation in PM2.5 concentration, respectively. The feature pairwise interactions account for 45.8% of ambient particle concentrations. The combination of trees and hedges in urban fabrics with longer building length than width could mitigate PM2.5 concentrations in residential areas. The integrated methods in this study could be extrapolated to future research to perform source identification analysis.