Deep learning for satellite data-driven assessment and forecasting of particulate pollution over South Korea

As the population grows, energy use increases, and long-range transport of pollutants from Chinese industrial activity evolves, air quality in Korea has become a problematic issue. As a result, there is strong demand for better air quality estimation in Korea, which we have addressed by developing a novel modeling framework based on Copula Bayesian Networks (CBN). CBN allows to combine data from multiple sources, including ground-level particulate matter air pollution (PM2.5 and PM10) from multiple monitoring stations, satellite observations of aerosol optical depth (AOD), important secondary aerosol precursors (e.g. ammonia), meteorological factors (pressure and wind patterns), and complex non-linear temporal and spatial scale dependencies inherent in these data. As many of these data originate on disparate spatial scales, they were remapped to a common grid using the Hierarchical Equal Area isoLatitude Pixelization (HEALPix).

Implemented on the Amazon Elastic Compute Cloud (Amazon EC2), our CBN approach predicted particulate matter air pollution with primary inputs of aerosol observations from the recently released V23 Multi-angle Imaging SpectroRadiometer (MISR) aerosol product, Geostationary Ocean Color Imager (GOCI), and ammonia observations from the Infrared Atmospheric Sounding Interferometer (IASI). Two-dimensional cloud motion vectors also showed to be important predictors of PM2.5 and PM10. All models were cross-validated and sensitivity analyses were conducted to address uncertainties in our model results.