Production of meteorological-spatial nitrogen dioxide (NO2) data using remote sensing and ground-based measurements

A modelling study, using open-source data sets, was conducted to increase the precision in determining the distribution of atmospheric nitrogen dioxide (NO2) pollution, due to the insufficiency of ground based point measurements in capturing the spatial distribution of the the air pollutants. In the study, the monthly mean concentration of atmospheric NO2 from (i) 15 ground-based Continuous Monitoring Centre (CMC) stations (Bahçelievler, Demetevler, Etimesgut, Keçiören-Sanatorium, Ostim, Sincan, Siteler, Törekent, Dissemination-Mamak, Dissemination-Çankaya, Polatlı, Çubuk, Ulus-Traffic, Sıhhiye, and Yaşamkent) in Ankara in 2021, these have the highest data quality (highest number of continuous hourly samples) among CMCs available, (ii) the NO2 satellite images from Sentinel5P Tropomi, cloud optical depth, (iii) TerraClimate minimum temperature, (iv) maximum temperature, (v) vapor pressure, (vi) vapor pressure deficit, (vii) wind speed, and (viii) NOAA/VIIRS's Average Light Brightness Value data were used to create a surface level NO2 prediction model for urban air quality monitoring using machine learning techniques. The statistical evaluation of the study was performed by using linear methods such as Multiple Linear Regression (MLR), Principal Component Regression (PCR), and Partial Least Squares Regression (PLSR). The direct linear relationship among variables examined via MLR, while, PCR and PLSR preferred and applied to reduce multi-dimentionality of correlated prediction variables. Ground-based monthly mean NO2 measurements, NO2 satellite observations, optical depth, minimum temperature, maximum temperature, vapor pressure deficit, wind speed, and average light brightness parameters were used in prediction model studies such as MLR, PCR, and LSR. The aim was to determine the model performance of each technique by statistically evaluating the satellite-based prediction parameters obtained via model and comparing with the real-time ground based measurements. As a result, MLR model has a weak correlation (or no significant correlation) between two datasets, while PCR and LSR model proposes a highly correlated datasets. In the study, the relationship effect was reduced, the percentage it explained increased, and it was determined that the error and correlation values had a negative effect. In the applications of the MLR, PCR, and PLSR methods, adjusted R-square values were obtained as 0.27, 0.48, and 0.49, respectively. The correlation increased in the second rank (at PCR and LSR) by eliminating the effect of multicollinearity. It was determined that the model generated with the PCR andPLSR machine learning methods potentially explains the atmopsheric NO2 measurements at surface level by 87% and 85%, respectively. .