The influence of neighborhood-level urban morphology on PM2.5 variation based on random forest regression

To improve the atmospheric environment by optimizing urban morphology, this study develops a random forest (RF) model to investigate the influence of urban morphology on PM2.5 variations via the relative importance of urban morphology and the nonlinear response relationship between urban morphology and PM2.5. Two indices—reduction range (C↓) and rate (C˅) of PM2.5 concentrations—are defined to evaluate the temporal variations of PM2.5. Results show that RF models are more accurate and perform better than multiple linear regression models, with R2 ranging from 0.861 to 0.936. Five out of nine urban morphological indicators have the most significant contribution to PM2.5 reduction. For each indicator, the nonlinear response relationship shows similar trends in general, despite of the difference at the higher pollution level. Building evenness index and water body area ratio have a similar response such that C↓ and C˅ sharply increase and tend to be stable when they reach at 0.05 and 8 %, respectively. With the increase in vegetated area ratio, the change of C↓ presents an inverted V-shape trend with the turning point of about 20 %; however, the change of C˅ greatly differs from the pollution level. A higher density of the low-rising buildings with one to three floors will lead to a small reduction rate but a greater reduction range of PM2.5. Floor area ratio values generally show a negative and nonlinear influence on C↓ and C˅. This study provides useful implications for planners and managers for PM2.5 reduction through neighborhood morphology optimization.