Quantifying the reduction in particulate matter by urban park green spaces using a random forest model: A case study...

Urban particulate matter (PM) pollution poses a serious threat to residents' health, and urban park green spaces (UPGSs) can play an important role in reducing the PM. Previous studies have highlighted how differences in plant traits, community structure, and green space configuration influence PM reduction. However, the extent to which meteorological factors, park built-up environment indicators, and other variables differentially drive PM reduction by UPGSs remains unclear. This study investigated 37 parks in Hohhot to assess the reduction rates (D_R) of 6 p.m. size fractions using summer daytime data. A multistage modelling framework was employed to compare the linear, generalised additive, and random forest (RF) models. RF was selected as the better-performing model through leave-one-park-out cross-validation (point-level R²: 0.19-0.58; park-level R²: 0.18-0.72). SHapley Additive exPlanations identified key driving factors, while the derivatives of partial dependence plots and individual conditional expectation plots quantified their marginal effects, revealing nonlinear relationships. The results showed that: (1) PM concentrations and D_R varied markedly across parks, exhibiting considerable spatial heterogeneity; (2) nine primary drivers were identified, and their comprehensive threshold ranges during summer daytime were determined: meteorological factors (air temperature: 22.08-33.26 °C, relative humidity: 29.71-67.17%, dew point temperature and wet bulb temperature: 11.47-24.38 °C, and wind speed: 0.14-1.12 m/s), park built-up environment indicators (distance from the city centre: 2.7-16.79 km), park spatial form indicators (perimeter-area ratio: 0.15-0.33, contiguity index: 0.24-0.48), and landscape indices (aggregation index: 78.46-86.85%). Beyond these ranges, the marginal mitigation effects plateaued. Therefore, strategically regulating key driving factors within their optimal ranges can effectively maximise D_R, thereby providing actionable guidance for enhancing the ecological health services of UPGSs.