?-Net: Superresolving SAR Tomographic Inversion via Deep Learning

Synthetic aperture radar tomography (TomoSAR) has been extensively employed in 3-D reconstruction dense urban areas using high-resolution SAR acquisitions. Compressive sensing (CS)-based algorithms are generally considered as the state-of-the art super-resolving TomoSAR, particular single look case. This superior performance comes at cost of extra computational burdens, because sparse reconstruction, which cannot be solved analytically, and we need to employ computationally expensive iterative solvers. In this article, propose a novel deep learning-based TomoSAR inversion approach, $\boldsymbol {\gamma }$ -Net, tackle challenge. -Net adopts advanced complex-valued learned shrinkage thresholding algorithm (CV-LISTA) mimic optimization step reconstruction. Simulations show height estimate from well-trained approaches Cramér-Rao lower bound (CRLB) while improving efficiency by one two orders magnitude comparing first-order CS-based methods. It also shows no degradation super-resolution power state-of-the-art second-order solvers, much more than Specifically, reaches 90% detection rate moderate cases 25 measurements 6 dB SNR. Moreover, simulation limited baselines demonstrates that proposed outperforms method fair margin. Test on real TanDEM-X data with just six interferograms high-quality high-density detected double scatterers.