Research on the impact of urban built environments on PM2.5 pollution based on machine learning methods

Abstract

Since PM_2.5 pollution poses a serious threat to the environment and health, understanding its interaction with the urban built environment (UBE) is essential for effective mitigation. To assess the impact of UBE on PM_2.5 pollution, this study quantitatively evaluates the relationship between them. First, given the limitation that current PM_2.5 concentration collection mainly relies on fixed monitoring stations, this study set up a taxi mobile monitoring system. Second, aiming at the deficiency of traditional extraction mostly based on remote sensing imagery, this study proposed a deep learning-based method to calculate the green and sky visibility index. Pearson's preliminary correlation analysis showed that climate factors were most correlated to changes in PM_2.5 concentration. Furthermore, the prediction effects of nine mainstream machine learning methods were compared. The results showed that (1) The overall prediction performance of summer (R² = 0.92) and autumn (R² = 0.93) outperformed the one of spring (R² = 0.88) and winter (R² = 0.86) seasons. (2) The Random Forest and LightGBM models obtained optimal predictions with R² of 0.907 and 0.916, respectively. (3) The complex nonlinear relationship between the UBE and PM_2.5 concentration needed to be captured by the Shapley additive explanations method. The findings suggested controlling the space enclosure index between 0.08 and 0.15, plot area ratio within 0.5, and building density within 0.2. This study provided a general analytical framework for understanding the diffusion mechanism of PM_2.5 concentrations and a theoretical basis for green urban design.