Gaussian Markov random fields improve ensemble predictions of daily 1 km PM2.5 and PM10 across France

Abstract

Understanding the health impacts of particulate matter (PM) requires spatiotemporally continuous exposure estimates. We developed a multi-stage ensemble model that estimates daily mean PM2.5 and PM10 at 1 km spatial resolution across France from 2000 to 2019. First, we alleviated the sparsity of PM2.5 monitors by imputing PM2.5 at more common PM10 monitors. We also imputed missing satellite aerosol optical depth (AOD) based on modelled AOD from atmospheric reanalyses. Next, we trained three base learners (mixed models, Gaussian Markov random fields, and random forests) to predict daily PM concentrations based on AOD, meteorology, and other variables. Finally, we generated ensemble predictions using a generalized additive model with spatiotemporally varying weights that exploit the strengths and weaknesses of each base learner. The Gaussian Markov random field dominated the ensemble, outperforming mixed models and random forests at most locations on most days. Rigorous cross-validation showed that the ensemble predictions were quite accurate, with mean absolute error (MAE) of 2.72 μg/m3 and R2 of 0.76 for PM2.5; PM10 MAE was 4.26 μg/m3 and R2 0.71. Our predictions are available to improve epidemiological studies of acute and chronic PM exposure in urban and rural France.