Synthetic Aperture Radar (SAR) imagery offers many ways of research in remote sensing. The SAR satellite sensors can provide high resolution images whatever the weather is, by day or by night. Among all the possible uses of radar images, we focus here on the way to obtain Digital Elevation Models (DEM) from pairs of stereoscopic SAR images: the radargrammetry. Radargrammetric application requires mastery of several steps: viewing geometry, image matching and geometric reconstruction solving 3D stereo intersection equations. The image matching method is an area correlation, coming from photogrammetric methods used in optical imagery. To improve performance during the matching step, it's interesting to implement speckle (particular multiplicative noise in radar images) filtering and to use a pyramid scheme in order to speed up the process and to reduce matching errors. Moreover, using epipolar geometry reduces the search area. In our study area with high relief, these methods can generate a quite good DEM but not enough accurate. The methods coming from photogrammetry are not sufficient to be directly applied to radar images. Specific methods are needed, so we propose a new approach using different correlation windows sizes: the multi-window approach. We particularly reduce the correlation windows along the range axis depending on the image range compression. This technique allows to make a reliable matching and to improve the reconstruction results accuracy.
