Taylor-Linearized ADMM Deep Unfolding Network With Classification Prior Guidance for Tomographic Inversion of Urban Areas

Abstract

Tomographic synthetic aperture radar (TomoSAR) is an essential technique for 3-D urban reconstruction. Traditional compressed sensing-based TomoSAR imaging methods provide super-resolution ability but suffer from high computational complexity and empirical parameter tuning. Recently, deep learning networks have improved the efficiency of TomoSAR reconstruction. However, the existing deep-learning-based TomoSAR reconstruction methods suffer from inversion inefficiency and insufficient prior guidance. This article proposes a Taylor-linearized alternating direction method of multipliers (ADMM) deep unfolding network with classification prior guidance, named TP-ADMM-Net, for tomographic inversion of urban areas. The primal-variable update is linearized to replace matrix inversion with an explicit update, and a residual compensation module is designed to reduce the approximation errors caused by linearization. In the auxiliary variable update, a scatterer classification prior feedback mechanism is developed to adjust the soft-thresholding parameters at different elevation grids. An outlier suppression strategy based on spatial neighborhood continuity is further introduced as a postprocessing step to reduce the isolated elevation outliers. Experiments based on simulated data show that TP-ADMM-Net achieves the lowest mean absolute error of 0.4323 m. This value is 36.33% lower than the second-best result. Experiments based on the real Fucheng-1 spaceborne dataset and the SARMV3D-1.0 airborne dataset demonstrate the improved reconstruction continuity and computational efficiency. Compared with the second-best result in each evaluated sub-region, TP-ADMM-Net achieves average reductions of 35.63% and 33.89% in mean total variation and mean local standard deviation, respectively, on the airborne dataset. The corresponding reductions on the spaceborne dataset are 83.87% and 98.70%, respectively.