A two-stage algorithm to estimate ground-level PM2.5 concentration levels in Madrid (Spain) from AOD satellite data and surface proxies

Poor air quality in urban areas is an important health risk; therefore, reducing population exposure to pollutants such as PM_2.5 is a major concern. Health assessments regarding this pollutant have typically relied on the measurements from urban networks of Air Quality Monitoring Stations (AQMS) to assess population exposure. The methods used for the spatial interpolation of observation often lacks a solid physical basis. Mesoscale air quality models provide high spatiotemporally resolved ground-level concentrations based on urban features, including the distribution of pollution sources; however, they are subject to significant uncertainty. In this work, a novel methodology to produce 1 km² resolution maps of ground-level PM_2.5 concentration for the Municipality of Madrid during 2015 is presented. Toward this end, different data sets including: meteorology, satellite observations of atmospheric optical depth (AOD) from MAIAC, emission data, population, land use, and vegetation land cover have been used. Subsequently, we applied extreme gradient boosting (XGBoost) machine learning algorithms in two steps to first fill gaps in the AOD field and then, estimate ground-level PM_2.5 concentration. The predictions of the so-called 2_step_XGBoost algorithm were compared with observations from the all available ground-level PM_2.5 concentration observations from the AQMS in Madrid obtaining a determination coefficient (r²) of 0.96, a RMSE of 1.5 μg/m³, and negligible bias. Additionally, we used a 10-fold cross validation to confirm the robustness of the algorithm and the independency of the dataset used for training (r² of 0.94 ± 0.01, RMSE of 0.40 ± 0.04 and MAE of 0.22 ± 0.02. These results highlight the reliability of this approach for future urban health analysis. In addition, we performed a Feature Importance (FI) analysis that revealed that 2_step_XGBoost identified the planetary boundary layer height (PBLH) as the most influential variable while AOD was found to have relatively low explanatory power, a result that may be contrasted in other case studies.