Towards Realistic TomoSAR Reconstruction Through Light Reconstruction and Enhancement

Tomographic synthetic aperture radar (tomoSAR) is an effective tool for large-area urban 3D mapping. TomoSAR reconstruction is fundamental to the quality of the mapping product. Recently, deep learning (DL) techniques have shown promising reconstruction precision compared to traditional methods. However, current work processes the data in a fiber-by-fiber manner to assemble the final 3D scene tensor, which may deteriorate the correlations existing within these fibers due to the spatial structure in the scene. In this work, we propose a new reconstruction neural network that features light reconstruction and enhanced computational processing. We use 2D slices as the basic processing unit. First, the rough reconstruction result is obtained through a light reconstruction that regularizes only the sparsity feature. Then, to enhance the spatial structure, we process the rough result with an enhancement stage through a deep neural network module. To harness correlations within the 3D scene tensor, we adopt an inter-slice and intra-slice strategy. The spatial structure within slices is explored first through a U-Net structure, while the structure between slices is further explored through a Long Short-Term Memory (LSTM) based module. Both simulated and measured data are used to evaluate the reconstruction neural network. Results reveal a high gain of spatial features, such as more complete building surfaces, compared to current methods.