Titre/Title : Imagerie interférentielle en radar à ouverture synthétique

Abstract :

Phase differences between two images are used in synthetic aperture radar interferometry (InSAR) to measure topography or ground motion with very high accuracy. However, this technique requires the removal of phase ambiguity
by the so-called process. Satellite data make this process quite difficult because of increased noise, decorrelated areas and discontinuities due to the radar sensor.

The goal of this thesis is to bridge the gap between raw interferograms and reliable information required for automatic phase unwrapping. After reviewing InSAR characteristics and the main techniques used in this field, we shall propose several algorithms adapted to interferometric data, which together constitute a robust processing method.

Our approach starts with an estimation process which measures the local two-dimensional frequency of the complex phase signal. A multi-scale spectral analysis algorithm provides fringe width and orientation, together with a measure of confidence in each estimate.

We first use these results in a filtering process meant to remove the wrapped phase noise. Our algorithm is adapted to the fringe pattern by performing a local compensation of the terrain slope or the field displacement. This enables us to average the right number of complex samples required to reduce the variance in the estimated phase, as a function of the correlation level.

Irreversible perturbations persist in areas where the geometric phase is inaccessable or under-sampled. Such areas are detected by a classification process using the previous measure of confidence and the correlation and amplitude images available along with interferograms. Fusion of the several measures characterizing the different perturbations takes into account contextual information through a markovien model adapted to the detection of thin structures linked to foreshortening effects.

The results of the above three processes allow us to unwrap the phase by local propagation techniques or weighted least squares global solutions. Their performances are evaluated when using the mask obtained by classification and the phase gradient measured, either by filtered phase differences, or directly by the local frequencies.

The proposed algorithms have been tested on interferometric products generated by CNES (the french space agency) from ERS-1 satellite images. Thus, the whole approach is validated on both topographic and differential real data with various characteristics which cover the most often encountered perturbations.