Modelling and prediction of daily, pan-European estimates of PM2.5 and PM10 based on Quantile Machine Learning applied to different mode Aerosol Optical Depth and reanalysis data

Among the different air pollutants, Particulate Matter (PM) poses a prominent threat to human health. In 2020, the exposure to PM2.5 (i.e. particles smaller than 2.5 micrometres in diameter) caused over 238,000 premature deaths in Europe, almost five times higher than the contribution from nitrogen dioxide, and ten times larger than ozone. Epidemiological studies for Europe generally rely on ground-level daily measurements to assess ambient PM exposures. However, the uneven distribution and discontinuous daily measurements of ground-level sites is a major constraint to develop large-scale, continental-wide epidemiological studies, biassing the results towards urban regions and areas with more ground-level sites.

In recent years, Aerosol Optical Depth (AOD) has increasingly become a useful alternative source of proxy data to estimate ground-level PM concentrations, because (i) AOD depicts the total column of aerosol in atmosphere, while PM depicts the surface aerosol, and (ii) its global spatiotemporal distribution can be easily obtained from satellites at high resolution. Despite the evident advantages, (i) the relationship between satellite AOD and PM is spatially heterogeneous, (ii) the number of missing data of satellite AOD is relatively high (up to 85% globally) due to cloudiness, and (iii)  the quality of measurements depends upon geographical factors like surface reflectivity. Europe is the one of the continents with lowest correlation between satellite AOD and PM concentration, so estimating PMs with satellite AOD in Europe becomes a great challenge. Furthermore, the components of AOD (fine and coarse-mode AOD, fAOD and cAOD respectively) are generally not available from satellite data. Thus, fewer studies used fAOD in the estimation of PM2.5, even when some studies found that fAOD is more highly associated with PM2.5.

Reanalysis data is another source to obtain available PM estimates,  (e.g., the PMs from Copernicus Atmosphere Monitoring Service Global Reanalysis (CAMSRA) and NASA’s Modern-Era Retrospective Analysis for Research and Applications v2 (MERRA-2)). However, they generally have lower resolution (on the order of 50-100 km) and have relatively large biases when it comes to the representation of surface pollution.

Here we downscaled and calibrated existing aerosol reanalysis, with the help of the AOD componental products (AOD and fAOD) generated in a previous study. To avoid the model  overfits in areas with dense monitoring sites (e.g., large cities), we used distance weighted loss functions (higher penalty weight on those places with fewer sites) to train the Quantile Machine Learning (QML) model. Then we predicted 18-year daily estimates and 95% predictive intervals for PM2.5 and PM10 at 10km resolution. In the model, we included atmosphere, land and ocean variables (e.g., boundary layer height, downward UV radiation, temperature, air pressure, humidity, cloud cover, local climate zone, leaf area index, surface reflectivity and road information). The results show that the out-of-sample r-squared (R2) of our PM2.5 and PM10 models is equal to 0.69 and 0.63, respectively, and largely outperform PM2.5 and PM10 estimates from CAMSRA (R2 = 0.25-0.35) and MERRA-2 (R2 = 0.22-0.33). Our approach provides more accurate PM estimates in Europe for the last 18 years, and opens new avenues for large-scale, high-resolution epidemiology studies.