CarbonNET: carbon dioxide retrieval from satellite using neural networks

In this work, we will show the potential of a nonlinear statistical regressor method based on a Deep Neural Network (DNN) scheme for retrieving XCO₂. Toward this objective, we set up a training exercise based on simulated IASI observations using the state-of-the-art radiative transfer mode (RTM) σ-IASI/F2N. A nine-year-long record from 2014 to 2022 of atmospheric state vectors using CAMS reanalysis dataset from ECMWF related to one day of each month at four synoptic hours (00-06-12-18 UTC) has been processed to capture typical seasonal and diurnal cycles, resulting in about 400,000 of IASI-L1 synthetic spectral radiances. In order to provide the regression scheme with the most representative information on the CO₂ signature, we implemented principal component analysis (PCA) of different regression features. Specifically, the PCA transform was applied to IASI band-1 (645-1210 cm^-1), which is most affected by CO₂ absorption, and to atmospheric temperature profiles. For IASI measurements the base of 90 principal components from the EUMETSAT IASI Level one Principal Component Compression (PCC) has been considered. Finally, different locations at various latitudes were selected to validate and evaluate the retrieval scheme's performance. In terms of validation, a set of real IASI soundings was matched with in situ measurements collected at Mauna Loa station, renowned as a background site with minimal regional impact. Preliminary findings demonstrate a high level of accuracy in extracting growth rate, trend, and seasonality from the predictions, showing a correlation greater than 0.9 with the in-situ data.