Estimation of near-surface NO2 based on TROPOMI and provincial control stations' data using machine learning

Abstract

Near-surface nitrogen dioxide NO2 concentration is an essential indicator for ambient air quality monitoring. In this study, the extreme gradient boosting (XGBoost) algorithm and random forest algorithm in machine learning are used to combine the TROPOspheric Monitoring Instrument(TROPOMI) high-resolution remote sensing images, meteorological and other auxiliary data with ground-level NO2 concentration monitoring data (including national and provincial control stations) to construct a dataset of estimation samples and conduct research on estimating the near-surface NO2 concentration on a grid with a spatial precision of 0.05° in Sichuan Province. According to the ten-fold cross-validation results of the test and training sets, the XGBoost model has better accuracy and generalization performance (R2=0.875, RMSE=4.774 ug/m3), In addition, SHapley Additive exPlanation(SHAP) was employed after its development, which showed that TROPOMI satellite data contributed the most to the near-surface NO2 estimation. By contrasting the results with the data set using only national control stations, we can see that after including the atmospheric monitoring data from provincial control stations, the estimated NO2 concentration near the ground is more consistent with the data distribution of ground monitoring stations. Moreover, the spatial distribution of concentration is more continuous and homogeneous, providing essential support for local governments to regulate the atmospheric environment precisely.