A satellite-based spatio-temporal machine learning model to reconstruct daily PM2.5 concentrations across Great Britain

Epidemiological studies on health effects of air pollution usually rely on measurements from monitors, which provide limited spatio-temporal coverage. Data from satellites, reanalysis and chemical transport models offer additional information used to reconstruct pollution concentrations at high spatio-temporal resolution. The aim of this study is to develop a multi-stage satellite-based machine learning model to estimate daily fine particulate matter (PM2.5) levels across Great Britain during 2003-2018. This high-resolution model consists of random forest (RF) algorithms applied in four stages. Stage-1 augmented monitor-PM2.5 series using co-located PM10 measures. Stage-2 imputed missing satellite aerosol optical depth observations using atmospheric reanalysis models. Stage-3 integrates the output from previous stages with spatial and spatio-temporal variables to build a prediction model for PM2.5. Stage-4 applied Stage-3 models to estimate daily PM2.5 concentrations over a 1-km grid. The RF architecture performed well in all stages, with results from Stage-3 showing an average cross-validated R2 of 0.788 and minimal bias. Spatial and temporal scale also performed well with R2 of 0.822 and 0.779, respectively. The high spatio-temporal resolution and relatively high precision allows this dataset (1.37 billion points) to be used in epidemiological analyses to assess health risks associated with both short- and long-term exposures to PM2.5.