Spatiotemporally resolved ambient particulate matter concentration by fusing observational data and ensemble chemical transport model simulations

Raw model data were found to be affected by a bias with a strong seasonal dependency: a large negative bias in winter and a small bias in the summer months. The data assimilation approach, embedded into a Bayesian hierarchical approach, was able to drastically reduce the bias. Furthermore, an advanced computational approach, based on the variational Bayes method coupled with the minimization of the Kullback Leibler divergence to approximate the optimal solution, made it possible to cost-effectively assimilate data throughout the period under consideration. By using stratified cross-validation to test the accuracy of our predictions, we found high out-of-sample R-2 ( = 0.83) and an average decrease of about two-thirds of the root mean square error.

In this work, we describe and implement a data assimilation approach for PM10 pollution data in Northern Italy. This was done by combining the best available information from observations and chemical transport models. Specifically, by (1) incorporating PM10 surface daily concentrations and model results from the CAMS (Copernicus Atmosphere Monitoring Service) ensemble; and (2) spreading the forecast corrections from the observation locations to the entire gridded domain covered by model forecasts by means of a data regularization approach. Results were verified against independent PM10 observations measured at 169 stations by local Environmental Protection Agencies. Twelve months of observations were matched in time and space, from January to December 2017, with air pollution model results. The studied domain encompassed the Po Valley, one of the most polluted areas in Europe, and that still does not meet the air quality criteria for the annual average concentration and the maximum number of exceedances allowed for the particulate matter.