Space remote sensing makes it possible to remotely measure the composition of the atmosphere thanks to its interaction with natural or artificial radiation. Indeed, the presence of molecules on the radiation path modifies the spectral content at frequencies characteristic of the various chemical elements. It is then possible to determine the concentration of different gases in an atmosphere column by comparing the measured spectral content with a radiative transfer model. Measurement accuracy is a critical component of mission performance.
The measurement instruments are high resolution spectrometers, whose spectral bands are centered around the characteristic frequencies of the gas to be observed. They comprise optical and electronic elements which can induce distortions and measurement errors. Superimposing on the distortions due to atmospheric gases, these instrument defects distort the measurement. It is thus fundamental to accurately characterize the instrument model and to take it into account correctly in the radiative transfer model.
The instruments are first characterized on the ground (e.g., thermal vacuum measurement campaigns), then at the beginning of their life in orbit (CALVAL phase), to validate the instrument and level 1 processing performance. Throughout their life in orbit, the quality of the products delivered to users is optimized by regular calibrations of instrument defects and by compensating for degradations due to ground / flight movements or aging.
Two research studies conducted in 2019 and 2021 with the French laboratory TeSA and supported by CNES have shown that it is possible to reconstruct accurately the Instrument Spectral Response Function (ISRF) by solving a linear inverse problem. This reconstruction is possible for a known input spectrum (available on ground or during flight calibrations) using the measurements. However, the ISRF does not account for radiometric imperfections (nonlinearity, straylight), which also affect the measurements. Consequently, it is important to enrich the estimated instrument model by taking into account at least the most impacting defects (or residues after calibration). Yet, such additional unknowns in the model may render the inverse problem nonlinear, complicating the analysis significantly.
The objectives of this PhD thesis are precisely to look for new methodologies for solving inverse problems with model imperfections, in order to improve the instrument model, with expected benefits for both ground characterizations and flight calibration.
Machine learning methods have recently succeeded in solving, with great precision, problems for which classical approaches had reached their limits. Schematically, machine learning aims at identifying the functional relationship between an input signal (e.g., input spectra, observable variables of an instrument) and an output signal (e.g., an ISRF, PSF, calibration parameters) from available examples (training data). The learnt relationship can be very complex and nonlinear. Moreover:
- The learnt relation is typically the inverse relation (solution of the inverse problem).- Precise knowledge about the model and its imperfections is not necessarily required (and learnt from data).- The learnt functional relationship is encoded by, e.g., a neural network, and can be applied at little cost to new observations.
The objective of this thesis is thus to propose and study new models that address the above current limitations and make it possible to, e.g.,:
- model and calibrate the imperfections of an instrument ,- estimate instrument parameters from incomplete observations,- compensate for uncertainty in the data or model ,- subsume temporal evolutions caused by movements, aging, seasonal effects .
In the blind inverse problem scenario (parameters / model partially unknown), this will require developing new inversion methods, using new learning strategies based on semi-supervised or unsupervised statistical learning. These will be applied to practical use cases provided by the CNES.
 Zimmerman et al. A machine learning calibration model using random forests to improve sensor performance for lower-cost air quality monitoring. Atmos. Meas. Tech. 11.1(2018).
 Jia et al. Point spread function modelling for wide-field small-aperture telescopes with a denoising autoencoder. Notices Royal Astron. Soc.493.1(2020).
 Wu et al. Sensor Drift Compensation Using Robust Classification Method.""Int. Conf. Neural Info. Proc.(2020).
For more Information, contact Directeur de thèse : email@example.com
about the topics and the co-financial partner (found by the lab !). Then, prepare a resumé, a recent transcript and a reference letter from your M2 supervisor/ engineering school director and you will be ready to apply online !
Message from PhD team
CNES will inform about the status of your application in mid-June. More details on CNES website : https://cnes.fr/en/web/CNES-en/10685-st-doctoral-grants.php