Estimate of XCO2 and Psurf from OCO-2 measurements using a Neural Network approach

Francois-Marie Breon,  CEA,  fmbreon@cea.fr (Presenter)
Leslie David,  CEA,  leslie.david@lsce.ipsl.fr
Frederic Chevallier,  CEA,  frederic.chevallier@lsce.ipsl.fr
Pierre Chatelanaz,  ENS Paris-Saclay,  pierre.chatelanaz@ens-paris-saclay.fr

We report on the progress made in the development of a Neural Network (NN) -based approach for processing OCO-2 radiance measurements. The first results showed a strong potential: after training against XCO2 values derived from an atmospheric transport model constrained by surface observations (CAMS), the NN applied to independent OCO-2 observations was in agreement with CAMS and with TCCON observations, with accuracies similar to those of the operational (ACOS) algorithm (David et al., 2021).
However, the NN retrievals did not show significant XCO2 enhancements downwind of high-emission cities or power plants. In addition, we have found that the NN approach is in fact capable of accurately estimating the observation date and latitude from its inputs (spectra and observation geometry), but not the longitude. We therefore hypothesize that this version of the NN uses the statistical information about XCO2 contained in the latitude and date rather than purely mining the input spectrum. However, the fact that the NN also works well for the surface pressure, for which a similar shortcut by date and latitude is not possible, is reassuring about the potential of the approach.
We found that the NN estimates the date from the information contained in the weak CO2 band. We have not yet pinpointed the feature that makes this possible, but we suspect that the NN estimates the observation date from the stratospheric CO2 concentration which has a signature in the spectra, and which increases regularly over time.
A new version of the NN approach was therefore developed using only the oxygen and strong-CO2 bands. Doing so, the NN no longer has indirect information on the date of the observation. The new version of the NN algorithm retrieves the XCO2 expected features, including the enhancements downwind of large emitters in the same way as the ACOS algorithm. We will describe the advantages and disadvantages of this new version of the NN retrieval method.

Poster: Poster_Breon__91_25.pdf 

Presentation Type: Poster

Session: 2.5a Retrieval algorithms and methods for inter-instrument and product Cal/Val

Session Date: Tuesday (6/15) 12:00 PM