Estimating surface sulfur dioxide concentrations from satellite data: Using chemical transport models vs. machine learning

Abstract. Sulfur dioxide (SO₂) is an important air pollutant that contributes to negative health effects, acid rain, and aerosol formation and growth. SO₂ has been measured using ground-based air quality monitoring networks, but the routine monitoring sites are predominantly placed in urban areas, leaving large gaps in the network in less populated locations. Previous studies have used chemical transport models (CTMs) or machine learning techniques to estimate surface SO₂ concentrations from satellite vertical column densities, but no direct comparisons between the methods have been made. In this study, we estimated surface SO₂ concentrations using Ozone Monitoring Instrument (OMI) retrievals over eastern China from 2015–2018 utilizing GEOS-Chem simulations and an extreme gradient boosting machine learning model. Compared to the in situ measurements, the SO₂ concentrations estimated from the CTM method had similar spatial distributions (r = 0.58) and intra- and interannual variations but were underestimated (slope = 0.24) with a relative percent error of ~75 % and had worsening performance over time. The machine learning method produced more accurate spatial distributions (r = 0.77) and temporal variations, a smaller discrepancy and bias (~30 %; slope = 0.69) and relatively stable performance over time. The machine learning method performed better than the GEOS-Chem method on smaller datasets and timescales with shorter temporal averaging periods. Ultimately, both methods were useful for estimating surface SO₂ concentrations since the CTM-based method does not rely on in situ monitoring and produced more reasonable spatial distributions than the machine learning method over areas without surface monitoring data.