Improving IFSAR phase unwrapping by early detection of non-interferometric features

This paper presents an original method base don the simultaneous analysis of phase, amplitude and coherence images for the detection of landscape features which do not provide topographic phase signal; for these surfaces, either phases or phase differences are not valid, respectively because of temporal changes and non-backscattering surfaces, or because of foreshortening and lay-over effects on mountain fore-slopes. A contextual classification based on Markov Random Fields is used to detect such perturbations through different classes corresponding to different feature shapes and characteristics. Classes describing thick structures are identified on the coherence and amplitude images available along with interferograms and on a measure of confidence derived from spectral analysis of the phase signal. The classical Potts model is well adapted to the regularization of such compact shape structures. Thin structures are enhanced by directional contrast operators dedicated to speckle SAR images. Associated directions allow a specific regularization scheme which preserves thin structures. The classification results are eventually used as a mask to enable automatic phase unwrapping performed either by path-following methods or by weighted least squares methods. Feature detection and unwrapping results are presented for various landscapes on interferometric data obtained from ERS-1 satellite over Ukraine and Switzerland regions.