Spatiotemporal Estimation and Analysis of PM2.5 Concentrations in Wuhan Utilizing Multisource Remote Sensing Data and NOx as Inputs for Machine Learning Models

Abstract

Atmospheric fine particulate matter (PM2.5) poses significant risks to both environmental and human health, highlighting the need for regional estimations and spatiotemporal analyses. While most studies have focused on large-scale areas, such as global or national levels, fewer studies addressed PM2.5 at the urban level. This study analyzed PM2.5 monitoring data from ground stations in Wuhan, collected between July 2018 and July 2023, integrating 1 km aerosol optical depth (AOD) products, Sentinel-5 NO2 column concentration data, nighttime light remote sensing, and ERA5 reanalysis meteorological data. Key innovations included selecting NO2 column concentration data, as NOx primarily exists as NO2, and using novel Sentinel-5P measurements rarely explored in related research. Three PM2.5 estimation models were developed: multiple linear regression (MLR), extreme gradient boosting (XGBoost), and random forest (RF). Evaluation results showed that all models achieved Pearson’s correlation coefficients (r) above 0.8, with the segmented RF-XGBoost model performing best, reaching an average relative error of 10.38%. Using this optimal model, monthly spatiotemporal maps of PM2.5 concentrations in Wuhan were generated. Key findings include 1) seasonal PM2.5 levels in Wuhan were lower in summer and higher in winter; 2) significant regional disparities in PM2.5 levels were observed, with persistently high pollution in areas such as Qing Shan; and 3) significant changes in PM2.5 levels before and after the COVID-19 pandemic, characterized by an overall decrease in concentrations from 2019 to 2020, followed by gradual increases in certain districts post-lockdown. This study provides valuable insights for urban-level PM2.5 estimation, supporting effective pollution control strategies.