Estimating PM2.5 with high-resolution 1-km AOD data and an improved machine learning model over Shenzhen, China

Abstract Studies on fine particulate matter with an aerodynamic diameter of 2.5 μm or smaller (PM2.5) are closely related to the atmospheric environment and human activities but are often limited by ground-level in situ observations. Satellite remote sensing techniques have been widely used to estimate the PM2.5 concentration over large areas where ground-monitoring sites are unavailable. However, satellite-retrieved aerosol optical depth (AOD) products usually feature a coarse resolution, which is insufficient for the estimation of the urban-scale PM2.5 concentration. We developed a new improved random forest (IRF) model based on machine learning and a newly released AOD product with a high resolution of 1-km, which could more effectively and accurately estimate the PM2.5 concentration over Shenzhen in the Guangdong-Hong Kong-Macao Greater Bay Area (GBA), China. Daily PM2.5 concentrations from 2016 to 2018 were estimated from ground-level PM2.5 and meteorological variable data. The popular linear regression model, geographically and temporally weighted regression (GTWR) model and random forest (RF) model without spatiotemporal information were employed for comparison and validation purposes through the 10-fold cross-validation (CV) approach. The IRF model attained an overall R2 value of 0.915 and a root mean square error (RMSE) value of 3.66 μg m−3. This suggests that the IRF model can estimate the urban PM2.5 concentration with a high spatial resolution at the daily, seasonal and annual scales, and the improved machine learning method is better than the linear model proposed by previous studies in terms of the estimation accuracy of the PM2.5 concentration. Generally, the IRF model coupled with AOD data with a 1-km resolution can significantly improve the calculation accuracy of the atmospheric PM2.5 concentration over coastal urban areas in the future.