Physics-aware TomoSAR Reconstruction with Deep-and-Shallow Hybrid Regularizations

Tomographic Synthetic Aperture Radar (TomoSAR) can produce three-dimensional images. Currently, imaging methods based on deep learning achieve high accuracy and efficiency but typically rely on supervised learning, which requires paired data. However, obtaining true scene values and highly reliable simulation data is challenging. To address this issue, this paper aims to achieve TomoSAR reconstruction using physics-aware unsupervised learning with deep-and-shallow hybrid regularizations. The proposed method generally employs the first-order gradient descent optimization algorithm for reconstruction. Here, the embedded proximal mapping steps are fused with both shallow and deep regularizations through the techniques of deep image prior and the plug-and-play denoiser, respectively. The deep regularizations are used in an unsupervised learning manner, with parameters gradually refined to optimal values during reconstruction without supervised learning. The proposed method is validated through both simulation and measured data imaging experiments. Results confirm that it is robust to noise interference while retrieving more scatterers in the scene, providing an alternative way to harness the strengths of deep learning while maintaining a physics-aware process.