Enhancing spatiotemporal coverage of satellite-derived high-resolution NO2 data with a super-resolution model

Abstract

Recent advancements in satellites, such as the TROPOspheric Monitoring Instrument (TROPOMI) have improved the spatial resolution of atmospheric chemical products, enabling more accurate estimation of atmospheric chemical concentrations. However, these satellites still face coverage limitations, posing challenges for monitoring atmospheric chemistry over large spatiotemporal scales. We propose a novel deep-learning model, GM-ESR-ResNet, which uses advanced super-resolution reconstruction for the coarse-resolution OMI data, supported by geo-meteorological covariates, to enhance the spatiotemporal coverage of high-resolution NO₂ columns from the original TROPOMI product. The results indicate that GM-ESR-ResNet significantly outperforms interpolation methods and traditional machine-learning models (which did not incorporate OMI NO₂ columns in prior studies) for spatiotemporal imputation of TROPOMI NO₂ data. The model enhances spatial coverage (imputation ratio = 30 %, correlation r = 0.82 vs. r = 0.42) with superior temporal transferability (MAPE of historical application 0.28 vs. 0.29). Additionally, the enhanced high-resolution NO₂ columns lead to more accurate surface NO₂ estimates than TROPOMI-based results, showing an increase in R² (0.856 vs. 0.871) and a 6 % reduction in RMSE. Furthermore, GM-ESR-ResNet, using OMI satellite data, successfully generates high-resolution NO₂ columns from 2015 to 2022, even in the absence of TROPOMI data. This enhancement improves the capture of strong spatial gradients in surface NO₂ concentrations and effectively addresses grid-to-point mismatch issues that previously led to substantial overestimations (by 11–30 %) in downwind and rural areas in the original coarse-resolution retrievals. Our study demonstrates the significant potential of super-resolution models in satellite retrievals, which is essential for aligning high-resolution column data with ground-based measurements.