A Deep Unfolding Network Algorithm for TomoSAR Building Object Height Inversion

TomoSAR building object height inversion is to utilize the two-dimensional radar image data obtained by a spacecraft such as a satellite that has sailed over a building several times to invert the position of the building scatterer distribution in the height direction, which is essentially a sparse reconstruction problem. The deep unfolding network algorithm ALISTA is a frontier algorithm for solving sparse reconstruction problems, which is a deep neural network algorithm obtained by combining the Iterative Soft Thresholding Algorithm (1STA) and Deep Learning, and possesses the advantages of both at the same time. The ALISTA-Polyak algorithm can be obtained by adding the Polyak heavy ball acceleration term to the iterative format of ALISTA, and this algorithm improves ALISTA at the level of first-order optimization algorithm. In this paper, on the basis of this algorithm, a densely connected network structure is introduced, and the ALISTA-Polyak algorithm is further improved at the level of deep neural network, and the DBALISTA-Polyak algorithm is obtained. The experimental results demonstrate that the inversion performance of the improved algorithm is superior to that of ALISTA and ALISTA-Polyak under the condition of restricted observation data.