01 Juillet – Thesis defense - Julien Boissy

09 h30 Amphi - laboratory IMS / buildint A31 (University of Bordeaux - Talence campus)

Microwave mapping and uncertainty quantification by Bayesian inversion-segmentation and Gibbs sampling.

An experimental set-up is dedicated to the microwave control of the homogeneity of electromagnetic properties, such as permeability and permittivity. By moving an antenna close to the material, it is possible to measure near-field reflection coefficients for various locations. One consider that the EM properties can be reduced to a surface impedance with possible discontinuities related to material inhomogeneities. The impedance can be linked to the reflection coefficients by a linearized forward model, derived from Maxwell's equations. It includes an additive term corresponding to various errors: measurement, linearization and modeling. Since the number of data available is small compared to the number of unknowns, the inversion of the observation model is tricky.
Therefore, we have developed a Bayesian inversion method that provides a segmented impedance map. A Gauss-Potts prior relies on a region division by a set of labels, modelled by a Potts field and on the impedance, modelled by Gaussian fields. Estimation is very tricky due to the large dimension and the non-standard form of the posterior. Consequently, we have turned towards a Gibbs algorithm that provides random samples of the aposteriori law, that are used to compute the estimate of the impedance map and labels, and to quantify uncertainties. The proposed method is validated on simulated data.

Event localization