Spatiotemporal analysis of PM2.5 estimated using machine learning over Greater Bangkok: Variability, trends, and persistence

Abstract

Abstract The estimation of surface PM2.5 over Greater Bangkok (GBK) was done using six individual machine learning models (random forest, adaptive boosting, gradient boosting, extreme gradient boosting, light gradient boosting, and cat boosting), and a stacked ensemble model (SEM) during the dry season (November–April) for 2018–2022. The predictor variables include aerosol optical depth (AOD) from the Himawari-8 satellite, a set of meteorological variables from ERA5_LAND and ERA5 reanalysis datasets, fire hotspots count and NDVI from MODIS, population density from WorldPop database, and the terrain elevation from USGS. Surface PM2.5 was collected for 37 air quality monitoring stations from the Pollution Control Department and Bangkok Meteorological Administration. A good agreement was found between Satellite AOD and AERONET AOD from two AERONET sites in GBK. Among individual models, light gradient boosting showed the best performance in estimating surface PM2.5 on both hourly and daily scales. The SEM outperformed all the individual models and hence was used for the estimation of PM2.5 for each grid in GBK for each hour. A higher risk of PM2.5 pollution in winter (November–February) as compared to summer (March–April) with a higher intensity in Bangkok province was evident from the spatiotemporal maps for both PM2.5 and its exposure intensity. The increasing trend in PM2.5 was reported over more than half of the area in GBK in winter and one-fifth of areas in summer. PM2.5 showed higher variability in winter as compared to summer which can be attributed to the episodical increase in PM2.5 concentration due to changes in meteorological condition suppressing dilution of PM2.5. The persistence analysis using the Hurst exponent suggested an overall higher persistence in PM2.5 during winter as compared to summer but opposite behaviors in nearby coastal regions. The results suggest the potential of using satellite data in combination with ML techniques to advance air quality monitoring from space over the data-scare regions in developing countries. A derived PM2.5 dataset and results of the study could support the formulation of effective air quality management strategies in GBK.