On the potential of a neural-network-based approach for estimating XCO<SUB>2</SUB> from OCO-2 measurements

International audienceIn David et al. (2021), we introduced a neural network (NN) approach for estimating the column-averaged dry-air mole fraction of CO2 (XCO2) and the surface pressure from the reflected solar spectra acquired by the OCO-2 instrument. The results indicated great potential for the technique as the comparison against both model estimates and independent TCCON measurements showed an accuracy and precision similar to or better than that of the operational ACOS (NASA's Atmospheric CO2 Observations from Space retrievals - ACOS) algorithm. Yet, subsequent analysis showed that the neural network estimate often mimics the training dataset and is unable to retrieve small-scale features such as CO2 plumes from industrial sites. Importantly, we found that, with the same inputs as those used to estimate XCO2 and surface pressure, the NN technique is able to estimate latitude and date with unexpected skill, i.e., with an error whose standard deviation is only 7° and 61 d, respectively. The information about the date mainly comes from the weak CO2 band, which is influenced by the well-mixed and increasing concentrations of CO2 in the stratosphere. The availability of such information in the measured spectrum may therefore allow the NN to exploit it rather than the direct CO2 imprint in the spectrum to estimate XCO2. Thus, our first version of the NN performed well mostly because the XCO2 fields used for the training were remarkably accurate, but it did not bring any added value. Further to this analysis, we designed a second version of the NN, excluding the weak CO2 band from the input. This new version has a different behavior as it does retrieve XCO2 enhancements downwind of emission hotspots, i.e., a feature that is not in the training dataset. The comparison against the reference Total Carbon Column Observing Network (TCCON) and the surface-air-sample-driven inversion of the Copernicus Atmosphere Monitoring Service (CAMS) remains very good, as in the first version of the NN. In addition, the difference with the CAMS model (also called innovation in a data assimilation context) for NASA Atmospheric CO2 Observations from Space (ACOS) and the NN estimates is correlated. These results confirm the potential of the NN approach for an operational processing of satellite observations aiming at the monitoring of CO2 concentrations and fluxes. The true information content of the neural network product remains to be properly evaluated, in particular regarding the respective input of the measured spectrum and the training dataset