Carbon monoxide (CO) is a product of incomplete combustion. It regulates the oxidising capacity of the troposphere, and it is one of the best tracers of wildfire and pollution plumes. As fires become more frequent and next-generation infrared sounders such as IASI-NG and MTG-IRS multiply the volume of data to process, the conventional processing chain becomes a bottleneck. In this chain, every interferogram is converted into a calibrated spectrum before an expensive full-physics retrieval is run. This thesis answers whether CO can instead be retrieved directly in the interferogram domain, and from how small a part of the interferogram.

The line-by-line radiative transfer and retrieval code LARA was extended to the interferogram domain. The extension includes the Fourier transform of modelled and observed spectra, an optical path difference (OPD) sampling adapted to the CO micro-window (2143-2181.5 cm-1), and a noise model derived from the instrument NEΔT. Since raw IASI interferograms are not distributed, we reconstruct them from Level-1C spectra, and the assumptions this process involves are stated explicitly. Tests on 14 941 scenes over the Atlantic Ocean show that neither the OPD sampling nor the assumed noise can be chosen freely. A factor four undersampling biases the CO column by +2.6% and removes 20% of the degrees of freedom for signal.

The first question is where the CO information lies in the interferogram, and whether it can be used without going back to a spectrum. An observing system simulation experiment (OSSE) was built with LARA on the TIGR archive, including 2311 atmospheric situations, crossed with eight CO profiles and seven surface temperatures, so that the true state is known exactly. A Jacobian analysis shows that the CO information is concentrated around the first harmonic of the ~4 cm-1 CO line spacing, near 0.25 cm OPD, and its multiples. Three algorithms of increasing complexity were then trained to map interferogram samples directly onto CO, H₂O and surface temperature. The best of them, a CNN-LSTM neural network using about thirty OPD samples, retrieves these quantities to within 15%, 20% and 0.3% respectively. It saves an estimated five hours of computation per hundred-pixel scene compared with a full-physics retrieval. However, it underestimates background CO levels by about 25%, exactly where a fire plume must be told apart from the background, and this low bias still prevents operational use.

The second question is whether the result holds on real observations with a physically based retrieval. The optimal-estimation retrieval of LARA was applied to IASI on Metop-B in three configurations that share the same forward model, a priori and noise, and differ only in the measurement vector, using spectral radiances, the full interferogram (0-2 cm OPD), and a partial interferogram (0.2-0.3 cm OPD). Over more than 11 500 soundings of a North Atlantic biomass-burning plume, the full interferogram and spectral retrievals agree to within 0.34% (r = 0.9998). Both agree with the operational FORLI product (biases of -0.4% and -0.7%, r = 0.975). Over nine years (2017-2025) at two TCCON stations, Park Falls and East Trout Lake, both agree with the ground-based reference within its stated uncertainty once the difference in vertical sensitivity is accounted for. To our knowledge, this is the first multi-year validation of an interferogram-domain product against a ground-based reference. The partial interferogram holds only 8 samples, a reduction of a factor 19 of the measurement vector (compared to the full interferogram), and loses about 27% of the DOFs. Yet it gives the smallest residual bias against TCCON (+0.73% and -0.56%).

Finally, an innovative instrument concept is introduced for a potential CO mission. Based on a static Fabry-Pérot array, it can only ever measure a partial interferogram. Hence, a preliminary thermal-infrared design adapted from the NanoCarb concept is proposed for CO and optimised with a Cramér-Rao lower bound analysis. The expected CO precision is about 10% for a single pixel and about 2% after along-track co-addition. This design opens the way to a compact CO sensor for persistent surveillance of wildfire plumes from a High-Altitude Platform System (HAPS).

Président du jury : M Jean Christophe Raut, Sorbonne Univeristé

Rapporteur : M Pierre Coheur, Université libre de Bruxelles 

Rapporteur : M Jean-Luc Attié, Université de Toulouse

Examinatrice : Mme Pascale Chelin, Université Paris-Est Créteil

Examinateur : M Hervé Herbin, Université de Lille

Examinateur : M Pierre-Yves Foucher, ONERA

Membre Invitée : Mme Carole Deniel, CNES

Directeur de thèse : M Sébastien Payan, Sorbonne Université

Co-directrice de thèse : Mme Laurence Croizé, ONERA