A data-driven method to retrieve XCO2 and XCH4 using artificial neural networks in preparation for the European Copernicus CO2 Monitoring Mission CO2M

Carbon dioxide (CO2) and methane (CH4) are the most important anthropogenic greenhouse gases because they are the main drivers of climate change. Monitoring their concentrations from space can help to detect and quantify anthropogenic emissions, supporting the mitigation efforts urgently needed to fulfill the Paris Agreement. Additionally, it can help to better understand the processes of the carbon cycle and thus allow better climate projections. These are key objectives of the European Copernicus CO2 Monitoring Mission CO2M, scheduled for launch in 2026, for which three retrieval algorithms are currently being developed and implemented in the EUMETSAT ground segment. These are so called conventional retrieval techniques that base on radiative transfer calculations. Despite shortcuts and approximations, the vast amount of satellite data makes them computationally expensive, requiring thousands of CPU cores. Although conventional retrieval methods base on physical principles, they typically require empirical data-driven methods to correct for biases in order to meet the demanding accuracy and precision requirements. The biases arise, e.g., from inaccuracies of the radiative transfer computations or unknown instrumental issues. Machine learning methods have the potential to combine both steps into a single data-driven retrieval algorithm, reducing the computational cost by several orders of magnitude. We used the radiative transfer model SCIATRAN to simulate two years (2015 and 2020) of sub-sampled realistic radiances of three instruments on board CO2M: the main instrument CO2I (CO2 imager), MAP (multi angle polarimeter), and CLIM (cloud imager). We use data from the first year of this data set to train artificial neural networks (ANNs) to retrieve XCO2 and XCH4 (the column-average dry-air mole fraction of atmospheric CO2 and CH4, respectively) plus related uncertainties and column averaging kernels. We will introduce a method which allows us to modify the training data making it representative for a wider range of atmospheric states. This ensures that the ANNs learn from the spectral signatures of CO2 and CH4 and that learning from spurious correlations is minimized. Despite the annual growth of CO2 and CH4, we will show that the ANNs trained with data from 2015 have almost the same quality when applied to data from 2020. We will analyze and compare the performance of different input vector settings, e.g., with and without MAP data and will discuss potential advantages or disadvantages of our ANN approach.