Near-Real-Time Prediction of PM2.5 Concentration over East Asia Using GOCI-II Aerosol Optical Properties and Machine Learning Model

Suspended particles with a diameter of 2.5 μm or less (PM2.5) are recognized to cause damaging effects on human health; higher asthma admissions in children and increased mortality rates are well known outcomes of short-term PM2.5 exposure (Fan et al., 2016; Kloog et al., 2013). While ground-level PM2.5 observations are conducted in most East Asian countries, the spatial limitation makes it difficult to assess the spatial and temporal variability of PM­2.5 over a wide area. As an alternative, aerosol optical depth (AOD) derived from satellite data is often implemented for estimating ground-level PM2.5 concentrations.

In this study, hourly ground-level PM2.5 concentration is predicted near-real-time using a Random Forest model over East Asia. Hourly aerosol optical products of GOCI-II (the 2nd Geostationary Ocean Color Imager) onboard a geostationary satellite GEO-KOMPSAT-2B, meteorological data, atmospheric chemical composition data and spatially and temporally inverse-weighted PM2.5 observation data are used as input variables. All input data were collocated to GOCI-II pixels with a spatial resolution of 2.5 km × 2.5 km. Before being fed into the model, relatively high-PM2.5 cases were oversampled to prevent underestimation. Prediction accuracy and diurnal variation of the simulated PM2.5 concentrations were evaluated. R2 peaked at local afternoon, while that of early morning decreased because fewer data points were available for evaluation, as GOCI-II observations are conducted only when the solar zenith angle is less than 70˚. Overall, the prediction resulted in similar values with ground-level PM2.5 observations, indicating a good model performance. The predicted PM2.5 concentration available near-real-time may be used as an indicator of health hazards on a timely basis, and can be applied to epidemiological studies when extended to longer timescales.