Estimating Hourly PM2.5 Concentrations from 2015 – 2018 in South Korea Using GOCI AOD

Abstract:

Air quality in South Korea has been rapidly deteriorating in recent years, as a result of both domestic sources and potential long-range transport from China. Ground PM_2.5 measurements in South Korea are restricted by fixed sites primarily located in urban regions, limiting complete spatial and temporal PM_2.5 coverage in the country. To address this limitation, AOD retrieved from the GOCI satellite was utilized in multiple random forest machine learning models to estimate hourly PM_2.5 concentrations. We developed 8 separate random forest machine learning models with time-varying meteorology and spatially fixed land information parameters to be included in the model. A 6 km modeling grid with 6,307 pixels was used to match the parameters to the GOCI retrievals. The average GOCI AOD coverage in the study domain was 48% across the study period. The 10-fold cross validated R2 ranged from 0.55 – 0.60 with RMSEs’ ranging between 9.48 – 12.44 µg/m³. The regression slope between observed and predicted hourly concentrations ranged between 1.1 – 1.2. Based on our models, we found that AOD retrieved from the GOCI satellite was ranked consistently as the most important predictor across all 8 models. This finding indicates that satellite measurements are an important source of data for exposure model development in South Korea. Overall, our hourly prediction models allow us to understand the spatiotemporal variations of PM_2.5 concentrations in South Korea that could not have been done by ground stations alone.