A TROPOMI-Corrected GEMS Product for Tropospheric NO2 Using Machine Learning to Correct Retrieval Errors

The GEMS geostationary satellite instrument launched in February 2020 is now providing continuous daytime hourly observations of tropospheric nitrogen dioxide (NO₂) columns over East Asia (5°S–45°N, 75°–145°E) with 7 × 8 km² pixel resolution at 37.5^o‑latitude. These data provide unique information to improve understanding of the sources, chemistry, and transport of nitrogen oxides (NO_x) with important implications for atmospheric chemistry and air quality. The early-generation GEMS retrievals display high biases compared to more mature satellite data (TROPOMI), but data for direct validation and bias correction are extremely limited. Here we apply a machine learning (ML) approach to correct the GEMS retrievals by reference to TROPOMI retrievals, fitting the Δ(GEMS-TROPOMI) differences to GEMS retrieval parameters as predictor variables including satellite geometry, cloud/aerosol properties, and surface reflectance. We train the ML model using collocated GEMS and TROPOMI NO₂ vertical column densities (VCDs) for 2022–2023. We then apply the trained ML model to the ensemble of GEMS data for 2022–2023, producing a blended GEMS+TROPOMI product with the coverage of GEMS and the precision of TROPOMI. We validate this blended product with ground-based observations from the Pandora spectrometers and use the ML model to identify the GEMS retrieval parameters associated with the largest discrepancies with TROPOMI.