Integrating low-cost air quality sensor networks with fixed and satellite monitoring systems to study ground-level PM_2.5

Airborne particulate matter (PM) mass concentrations measured by conventional monitoring stations are reliable data sources for air quality communication, pollution mapping, and exposure estimation. A high spatiotemporal density of monitoring stations can provide a better understanding of PM transport on a regional and global scale and help to reduce exposure misclassification leading to a better assessment of the health impacts associated with PM exposure. However, due to the cost and operational complexities, only a limited number of such PM monitors can be deployed. Apart from conventional measurements, PM mass concentration can be estimated from aerosol optical depth (AOD) data observations by using empirical, semi-empirical or modeling methods, but these datasets are usually compromised by weather conditions and lack of knowledge of aerosol properties. In addition to the above methods, a network of low-cost PM sensors is also a promising approach to increase the measurement density. In this study, we propose an integration of information from multiple measurement approaches. We demonstrate this approach by synergizing the data from 75 monitoring stations, 2,363 AirBox low-cost sensors (the amount of data entries is ∼10 million), and the Terra remote sensing satellite to estimate surface concentrations of PM for Taiwan Main Island during July 14 2018 to Oct 31 in 2018. A machine learning method selects the useful data from the low-cost sensor datasets, and the ordinary Kriging method is used to create a visual daily PM distribution map. The integration of datasets can enhance the overall data quantity and quality, leading to more accurate pollution maps with greater detail. The maps created from these three data sources demonstrate an approximate 30-fold synergistic improvement in the spatial resolution of PM mapping. The Root Mean Square Error (RMSE) of the predicted maps was analyzed through leave-one-out cross-validation, ten-fold cross-validation, and standard data validation. It shows that including low-cost PM sensor data brings in greater detail and largely enhances the spatial distribution while maintaining the pollution mapping characteristics. The approach described here will greatly assist the validation of PM transport models and enhance the accuracy of exposure estimations in future studies.