Deep learning-based downscaling of tropospheric nitrogen dioxide using ground-level and satellite observations

Air quality is one of the major issues within an urban area that affect people's living environment and health conditions. Existing observations are not adequate to provide a spatiotemporally comprehensive air quality information for vulnerable populations to plan ahead. Launched in 2017, TROPOspheric Monitoring Instrument (TROPOMI) provides a high spatial resolution (~5 km) tropospheric air quality measurement that captures the spatial variability of air pollution, but still limited by its daily overpass in the temporal dimension and relatively short historical records. Integrating with the hourly available AirNOW observations by ground-level discrete stations, we proposed and compared two deep learning methods that learn the relationship between the ground-level nitrogen dioxide (NO₂) observation from AirNOW and the tropospheric NO₂ column density from TROPOMI to downscale the daily NO₂ to an hourly resolution. The input predictors include the locations of AirNOW stations, AirNOW NO₂ observations, boundary layer height, other meteorological status, elevation, major roads, and power plants. The learned relationship can be used to produce NO₂ emission estimates at the sub-urban scale on an hourly basis. The two methods include 1) an integrated method between inverse weighted distance and a feed forward neural network (IDW + DNN), and 2) a deep matrix network (DMN) that maps the discrete AirNOW observations directly to the distribution of TROPOMI observations. We further compared the accuracies of both models using different configurations of input predictors and validated their average Root Mean Squared Error (RMSE), average Mean Absolute Error (MAE) and the spatial distribution of errors. Results show that DMN generates more reliable NO₂ estimates and captures a better spatial distribution of NO₂ concentrations than the IDW + DNN model.