Near-real-time hourly PM2.5 prediction over East Asia using geostationary satellite products and machine learning

Abstract

Airborne ground-level particulate matter of size <2.5 μm (PM2.5) is a potential health hazard, even with short-term exposure. To quantify the amount of ambient PM2.5, there have been many advances in the use of satellite observations in estimating PM2.5 concentrations over broader areas. Although hindcasts have successfully estimated PM2.5 concentrations at locations with no ground-based observations, nowcasts are crucial in preventing potential exposure to PM2.5 in real time. In this study, a near-real-time hourly PM2.5 prediction framework was designed using a geostationary earth orbit satellite, Geostationary Ocean Color Imager-II (GOCI-II) aerosol products, and random forest (RF) models to yield hourly PM2.5 predictions over East Asia 10 times daily at 2.5 km resolution. The framework includes an oversampling stage, which increases the number of infrequent PM2.5 occurrences. The algorithm was tested over a one-year period, during which predictions and observations were strongly correlated (R2 = 0.79) with a relative root mean square error of 41.99%. The results were comparable to or better than those of previous hindcast models. Uncertainties in prediction stemmed mainly from the number of valid predictions and the absolute volume of training samples. Due to the high temporal and spatial resolution, predicted PM2.5 concentrations successfully filled in regions with no ground-based stations. The products also identified concentration gradients within sub-districts, and their temporal changes during daytime. Prediction with a model with a training dataset from an hour ago still produced robust results. The products may be used in releasing health hazard alerts through real-time monitoring of air quality. © 2024