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 information for vulnerable populations plan ahead. Launched in 2017, TROPOspheric Monitoring Instrument (TROPOMI) provides high spatial resolution (~5 km) tropospheric measurement captures variability pollution, but still limited by its daily overpass temporal dimension relatively short historical records. Integrating with hourly available AirNOW ground-level discrete stations, we proposed compared two deep learning methods learn relationship between nitrogen dioxide (NO2) observation from NO2 column density TROPOMI downscale resolution. The input predictors include locations observations, boundary layer height, other meteorological status, elevation, roads, power plants. learned can be used produce emission estimates at sub-urban scale on basis. 1) integrated method inverse weighted distance feed forward neural network (IDW + DNN), 2) matrix (DMN) maps directly distribution observations. We further accuracies both models using different configurations validated their average Root Mean Squared Error (RMSE), Absolute (MAE) errors. Results show DMN generates more reliable better concentrations than IDW DNN model.