A data fusion uncertainty-enabled method to map street-scale hourly NO2: a case study in Barcelona

Considering that air pollution is the leading global environmental risk factor according to the WHO,  characterizing NO2 levels holds crucial significance, particularly in heavily trafficked urban areas where NO2 legal limits and health guidelines are frequently exceeded. Obtaining accurate and comprehensive NO2 datasets on a city level is especially challenging due to the inherent uncertainties associated with urban air quality models, and the scarcity of air quality monitoring stations. An alternative method to describe NO2 levels involves developing short-term experimental campaigns using indicative measurements, although they report period-averaged results and do not have full spatial coverage.  Taking advantage of the three mentioned approaches,  this work proposes a data-fusion method that combines i) near-real-time hourly observations obtained from the official air quality monitoring network, ii) the output of an urban air quality model (CALIOPE-Urban) that operates at high spatial (up to 20m x 20m) and temporal (hourly) resolutions, and iii) a microscale Land-Use-Regression (LUR) model based on machine learning. The microscale-LUR model includes different urban datasets such as traffic flow or average building density and two NO2 experimental campaigns.  While the hourly observations enable the temporal variability adjustment in the dispersion model, the microscale-LUR model provides additional insights into the spatial characteristics of NO2 distribution. Our data-fusion approach was implemented on an hourly basis over the metropolitan area of Barcelona in 2019. Besides the bias-corrected NO2 hourly maps, this method also computes the uncertainty associated with the variance of the estimated error during the correction process. By integrating both corrected NO2 values and their associated uncertainty, it produces maps that show the probability of exceeding the hourly 200 µg/m3 and the annual 40 µg/m3 NO2 legal thresholds over Barcelona.  Cross-validated results at the monitoring stations demonstrate that the spatial bias correction increases the correlation coefficient (r) by +46 % and decreases the root mean square error (RMSE) by −48 %, compared to the model output. This research emphasizes the importance of highly detailed spatial data within data-fusion techniques, enhancing the accuracy of predicting exceedances at the street level.