Methane profiling retrieval from IASI: a deep learning inversion approach based on feed-forward neural networks

In this work, a nonlinear statistical regressor method based on deep learning feed-forward neural network (NN) for the retrieval of atmospheric CH4 is proposed. The methodology has been trained and validated on a simulated dataset of observations by the processing of the Monitoring Atmospheric Composition and Climate (MACC) Reanalysis dataset with the state-of-the-art transfer model (RTM) sigma-IASI-as. Global data related to one day of the 12 months of 2012 and four synoptic hours (00-06-12-18 UTC) have been processed to catch typical seasonal and diurnal cycles, corresponding to a fairly large number (168.000) of simulated IASI-L1 spectral radiances. CH4 profiles have been predicted on 60 pressure layers. A regression framework based on the principal components analysis (PCA) of the IASI radiances and CH4 profiles has been implemented. The choice of the number of principal components has been addressed by the study of their eigenvalues, to filter redundant information from IASI channels and extract the most significant information from the CH4 profiles. The analysis of the NN retrieval, shows agreement with the reference MACC CH4 contents, allowing to obtain unbiased profile estimates, with accuracy on the total content of about 1.55%. The same accuracy has been obtained for the tropospheric column while for the stratosphere atmospheric column the accuracy is about 3%. Finally, an additional analysis of the CH4 total content shows a correlation between the reference and predicted values of about 0.97.