Leveraging wide area XCO2 deep learning in estimating urban CO2 emissions from space

Urban areas account for more than 70% of fossil fuel carbon dioxide (CO₂) emissions worldwide. Recent (OCO-3 released in 2019) and forthcoming (CO2M, TANSAT-2, and GOSAT-GW) greenhouse gas satellites can observe wide area column average dry air mole fraction of carbon dioxide (XCO₂) of entire urban areas. Although top-down urban emission monitoring has improved in terms of spatial coverage and frequency, the challenge remains in how to utilize space-based observations to perform accurate inversion of source area’s emission. The high uncertainty mainly arises from XCO₂ observations’ low signal-to-noise ratio due to non-anthropogenic fluxes and missing data due to sophisticated atmospheric conditions. 

To achieve accurate urban emission estimation from space, we propose a deep learning (DL) framework which can intelligently capture XCO₂ patterns from wide area XCO₂ observations. The synthetic CO2M dataset serves as model pre-training materials for its ideal XCO₂ observations given by chemical transport model. Transformer is selected as the architecture of DL model for its ability to model global dependency across wide area observations. The proposed model has been validated on the Berlin city’s synthetic CO2M dataset and OCO-3 snapshot area map (SAM) mode observations. In both cases, the pre-trained DL model effectively interpolated missing XCO₂ values throughout the XCO₂ snapshot, and showed outperformance on urban plume signal identification compared to conventional algorithms. Furthermore, by incorporating DL model’s prediction results with inversion methods, we performed emission estimates for Berlin city on synthetic CO2M data and multiple cities globally on OCO-3 SAMs. Our top-down emission estimation results showed high consistency with prior bottom-up inventories. This study provides valuable insights into advancing intelligent methodologies for urban emission inversion from wide area satellite observations.