Development of ground-level NO2 models in Vietnam using machine learning and satellite observations with ancillary data

Abstract

In this study, the aim was to create daily ground-level NO2 maps for Vietnam spanning from 2019 to 2021. To achieve this, various machine learning models (including the Mixed Effect Model, Neural Network, and LightGBM) were utilized to process satellite NO2 tropospheric columns from Ozone Monitoring Instrument (OMI) and TROPOMI, as well as meteorological and land use maps and ground measurement NO2 data. The LightGBM model was found to be the most effective, producing results with a Pearson r of 0.77, RMSE of 7.93 μg/m³, and Mean Relative Error (MRE) of 42.6% compared to ground truth measurements. The annual average NO2 maps from 2019–2021 obtained by the LightGBM model for Vietnam were compared to a global product and ground stations, and it was found to have superior quality with Pearson r of 0.95, RMSE of 2.27 μg/m³, MRE of 9.79%, based on 81 samples.