Estimating surface nitrogen dioxide and ozone concentrations using satellite-based and numerical model-based data

Prolonged exposure to high concentrations of nitrogen dioxide (NO₂) and ozone (O₃) at ground level could be harmful to human health. In-situ air pollutant concentration data observed from ground monitoring stations are limited in providing spatially continuous information. Since there are only a few stations installed above the sea, it is difficult to monitor the concentrations of air pollutants over the sea. In this study, machine learning-based models were developed to estimate ground-level NO₂ and O₃ concentrations using satellite-based remote sensing data and model-based meteorological and emission data over East Asia during 2015-2017, to overcome such limitations. NO₂ and O₃ vertical column density products from the Aura Ozone Monitoring Instrument (OMI) were used as essential predictors to estimate NO₂ and O₃ concentrations. Missing pixels of OMI products due to row anomalies were filled using a temporal convolution approach to generate the spatiotemporally continuous distribution of NO₂ and O₃ concentrations. In order to estimate the air pollutant concentrations in both land and ocean, specific values were assigned to the ocean for land-only variables. Random forest (RF) was used to develop the estimation models for NO₂ and O₃ concentrations. The RF-based models showed the results with R² values of 0.72 and 0.75, and RMSEs of 6.24 ppb and 10.56 ppb for NO₂ and O₃, respectively. The estimated results over the ocean were validated using coastal stations that are located within a 1 km distance from the coast. Compared to the model without land-only variables, the models using all variables had slightly better results. The satellite-based NO₂ and O₃ vertical column density were identified as significant variables in both models. Besides, urban land cover ratio, wind-related variables such as wind vectors, and stacked maximum wind speed had relatively high variable importance. The spatial variation of NO₂ and seasonal variation of O₃ were well shown in the estimated spatiotemporal distribution.