Sujet : « Traitement du signal pour les données spectrales de Mars ».
- Directeur de thèse : Frédéric Schmidt & co-encadrant : Matthieu Kowalsi (Labo L2S, SUPELEC, CNRS).
- Financement : Centre of Data Science, LIDEX Université Paris Saclay-COFUND.
The Planetary Fourier Spectrometer (PFS) instrument onboard the Mars Express
mission (European Space Agency, ESA) recorded a very large dataset of spectra,
thanks to 10 years of orbital observation around the Red Planet, at various places, local time and season, with unprecedented spectral resolution. Among other new detection, the PFS provided a very debated seasonal detection of atmospheric methane of several parts by billions. If true, this detection implies a large amount of methane released at present time, but also imply that methane is destructed very efficiently. Several candidate processes are proposed, including rocks alteration or life. This question has been considered at the top level by the ESA at the point to send the ExoMars Trace Gaz Orbiter mission (planed to launch in 2016 and arrive in 2017) to unravel this mystery. The main objective of this PhD program will be to correct the major limitation of the PFS instrument to have access to local atmospheric properties : the effect of microvibrations that can be corrected by blind deconvolution using complex variables.
Among blind inverse problems, blind deconvolution is one of the much harder.
Indeed, traditional mathematical priors such as sparsity inside dictionary of elementary atoms do not lead to satisfactory solution. However, a lot of advances have been ade, especially when strong priors can be made on the convolution filter, such as in speech dereverberation. Even if the transfer between the signal processing community and some specific applied domain is successful, the PFS instrument has not benefited yet of such advanced signal processing techniques. We plan to adapt and develop new deconvolution tools, and propose them as a end-user toolbox. Especially we will focus on the implementation on very parallelized architecture (GPU/Xeon Phi) to process very large dataset.