Spatio-temporal modeling of satellite-observed CO2 columns in China using deep learning

Satellite measurements of the column-averaged dry air mole fraction of atmospheric carbon dioxide (XCO2) play a crucial role in monitoring CO2 emissions and sinks. However, the current limitations of satellite observations, including sparse sampling, narrow swath coverage, and data gaps caused by factors like clouds, significantly hinder their ability to accurately capture local-scale CO2 sources and sinks. This study introduces an innovative data-driven approach based on deep learning, which takes into consideration both spatial and temporal variations, to map XCO2 using observations from multiple satellites. By leveraging advanced deep learning techniques like conventional neural network (CNN), long short-term memory network (LSTM), channel-spatial attention, and artificial neural network (ANN) in the model training, this approach not only incorporates spatiotemporal variations of XCO2 but also integrates information from related terrestrial, anthropogenic, and meteorological variables. The results demonstrate a notable improvement in the predictive capability of the approach. An important advancement over previous studies is that this approach breaks away from the conventional practice of using model-generated XCO2 simulations for both training and validation. Monthly deep-learning XCO2 (DL-XCO2) of China from 2014 to 2022 was generated with a spatial resolution of 0.25° based on satellite retrievals from GOSAT and OCO-2/3. Cross-validation results show an average prediction bias of −0.16 ppm. Additionally, DL-XCO2 exhibits high precision when compared to two TCCON stations, with errors of 0.93 and 1.29 ppm in Hefei and Xianghe, respectively. Ultimately, the DL-XCO2 data effectively capture urban CO2 variations, showcasing the potential in accurately characterizing fine-scale CO2 sources and sinks.