Conference Agenda

Synthetic Aperture Radar (SAR) tomography (TomoSAR) exploits multiple baseline acquisitions that are coherently combined to estimate a three-dimensional reflectivity distribution of the scene. Classical reconstruction methods, however, strongly depend on missions with a large number of tracks to achieve high-quality tomograms in terms of resolution and effective ambiguity suppression. This challenge is particularly pronounced in forested regions, where scattering contributions from vegetation volume and ground surfaces overlap along the reconstruction direction, increasing the dependence of the tomographic result on the stack size.

On the other hand, deep learning methods have gained increasing attention in remote sensing due to their ability to learn representative features directly from data. Within this framework, we propose to mitigate the limitations imposed by a restricted number of acquisitions by adopting a deep neural network based on an encoder–decoder architecture.

The proposed deep learning method is evaluated using L-band tomographic data acquired by the German Aerospace Center (DLR) F-SAR system over the Traunstein temperate forest site in Bavaria, southern Germany, in March 2017. The quality of the resulting tomograms is assessed using quantitative similarity measures by comparing them with those obtained using classical reconstruction techniques, such as the Minimum Variance Distortionless Response (MVDR) spectral estimator, also known as Capon.