Deep learning in SAR tomography

Tomographic SAR (TomoSAR) processing uses multi-baseline (MB) interferometric acquisitions and multi-polarization (MP) channels for reconstructing fully three dimensional reflectivity profiles of the observed ground scene. Due to the typical availability of few interferometric images with an uneven baseline distribution, coupled with super-resolution requirement, the problem to be solved for tomographic reconstruction is ill conditioned. Many regularized tomographic methods have been developed. However, these methods generally suffer from high computational overhead, presence of artifacts a low number of detected scatterers when the signal model deviates from the reference one. Moreover, many methods are based on the manual set of the algorithms parameters used to solve the optimization problem. Thus, several deep learning (DL) techniques have been developed for scene recovery. In this chapter, we briefly present some DL-based methods applied for TomoSAR reconstruction, distinguishing the case of urban scenes and forested scenes.