Priscilla Adong

Exposure to air pollution is one of the greatest environmental health risks globally with harmful effects on human health and the ecosystem [1, 2]. Among the key pollutants of concern is particulate matter (PM) including PM10 and PM2.5 characterised by aerodynamic diameters less than 10 μm and 2.5 μm, respectively [1]. In order to effectively manage these pollutants and evaluate the efficiency of pollution management strategies, it is essential to determine the pollutant levels and their variations in different environments over time. This is traditionally done through setting up ground monitoring networks. However, ground-based monitoring is very expensive to set up and maintain. In addition, the sparse coverage of ground-based monitors in low-resource settings is inadequate to quantify air pollution over large areas. On the other hand, satellites can gather data for almost all regions globally, thereby addressing the inadequate spatial coverage issues encountered with ground-based observations. Satellite observations have become increasingly prominent for tracking PM levels using satellite-derived aerosol optical depth (AOD) [2]. This technique provides a more comprehensive spatial overview of pollutant concentrations. Nonetheless, satellite measurements do not provide information regarding the vertical profile of pollutant concentrations, necessitating the use of estimation techniques to infer near-surface concentrations. Machine learning methods allow us to efficiently exploit the growing volume of satellite data available and accurately estimate near-surface concentrations. Dey et al [3] explored the use of convolutional neural networks to estimate near-surface concentrations for multiple trace gasses including nitrogen dioxide (NO2), sulfur dioxide (SO2), and ozone (O3). In this study, we will investigate the performance of various machine learning methods to estimate near-surface concentrations for PM.