Advantages of Polarimetry and Interferometry for Semantic Segmentation of Urban SAR Images with Consideration of the Layover

In order to make the most of SAR and its different modalities (PolSAR, InSAR, PolInSAR) for urban area monitoring, a semantic segmentation of dualpolarimetric TanDEM-X images into ground and above- ground classes is carried out. To account for the layover, a pixel may be assigned to both classes at once. Ground truth labels are constructed by projecting 3D LiDAR point clouds onto the azimuth/range reference frame of the SAR image. The impact of PolSAR and InSAR is studied through a comparison to models trained on a single polarization SAR image. PolSAR data are found to be beneficial, except when training and testing on different polarimetric couples. InSAR data are highly beneficial if their ambiguity height is small, and detrimental if it is much larger than the typical building height.