Skip to Main content Skip to Navigation

Assimilation de données pour les problèmes non-Gaussiens : méthodologie et applications à la biogéochimie marine

Abstract : Data assimilation for Geosciences is a discipline seeking to improve our knowledge of a physical system based on the information from numerical models simulating this system and the information from the measures observing this system. The data assimilation methods traditionally used (eg the 4DVAR or the ensemble Kalman filters) are based on assumptions of Gaussianity of the probabilities involved and linearity of the models. With the increasing complexity of models and observation networks, these assumptions are increasingly unjustified and therefore penalizing. This complexity is particularly strong in oceanography coupled with marine biogeochemistry.The objectives of this thesis are to understand the appearance of non Gaussianity in an estimation problem, to think out a data assimilation method adapted to highly non Gaussian problems and, in the coupling of ocean dynamics and marine biogeochemistry, to explore the relevance of the use of non Gaussian methods.At first, a methodological study is conducted. This study, supported by illustrations with the three variable Lorenz model, allows to highlight the limitations of traditional methods when facing non Gaussian problems. This study led to the development of a fully non Gaussian data assimilation filter : the Multivariate Rank Histogram Filter (MRHF).It is shown that the MRHF is efficient in highly non Gaussian regimes (including in a bimodal regime) for a relatively small number of members.Secondly, a numerical study is conducted. This study is conducted with twin experiments based on a 1D vertical model, ModECOGeL, coupling dynamics and biogeochemistry in the Ligurian Sea. We simulate different observation networks combining in situ profiles and satellite data. Several data assimilation methods are then compared using advanced ensemble evaluation diagnoses.Our experiments show the impact of observation networks and controled variables on the degree of non Gaussianity in an estimation problem. The control of the dynamic part of the model by observations of the dynamics at different frequencies is a quasi Gaussian problem, which a least squared filter such as the Ensemble Transform Kalman Filter solves well. In contrast, for the same observations, the control of biogeochemistry proves to be a non Gaussian problem and requires the use of a non Gaussian filter. Finally, it is shown that assimilation of ocean color data, for the joint control of the dynamic and the biogeochemistry, is improved by methods adapted for non Gaussianities such as the Anamorphosed Ensemble Kalman Filter. In addition, increasing the ocean color observation frequency makes unavoidable the use of fundamentally non Gaussian filters such as the MRHF.
Complete list of metadata

Cited literature [182 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Wednesday, April 27, 2016 - 2:52:19 PM
Last modification on : Thursday, November 19, 2020 - 3:54:27 PM
Long-term archiving on: : Thursday, July 28, 2016 - 10:52:36 AM


Version validated by the jury (STAR)


  • HAL Id : tel-01308288, version 1



Sammy Metref. Assimilation de données pour les problèmes non-Gaussiens : méthodologie et applications à la biogéochimie marine. Océan, Atmosphère. Université Grenoble Alpes, 2015. Français. ⟨NNT : 2015GREAU019⟩. ⟨tel-01308288⟩



Record views


Files downloads