Research On Multi-Baseline Tomography SAR Compressed Sensing Imaging Method Based On Deep Learning

Multibaseline Tomographic Synthetic Aperture Radar(MB-Tomo SAR)is a 3D SAR technology,which achieves 3D imaging by multi-trajectory observation of the platform without changing the current SAR system,and has practical engineering significance.In practice,the resolution and imaging quality of the perpendicular line of sight(PLOS)direction are low due to the small number of observations along the PLOS direction,ununiformly distributed and short spatial baseline.In order to improve the resolution and reduce the sidelobes,researchers have applied the compressed sensing(CS)algorithms to achieve super-resolution Tomo SAR imaging.However,existing imaging algorithms based on CS require multiple iterations and manual tuning,resulting in low computational efficiency and poor accuracy.In view of the advantages of deep learning SAR imaging methods in improving imaging accuracy and imaging efficiency at this stage,this thesis introduces deep learning technology into Tomo SAR imaging and conducts research on high-resolution Tomo SAR compressive sensing imaging methods.The main work and innovation points of this thesis are as follows.1.A MB-Tomo SAR imaging method based on Ada Tomo-Net is studied.The MBTomo SAR imaging method based on LFISTA has higher reconstruction accuracy and efficiency compared with the traditional CS algorithms,however,the reconstruction accuracy of the method decreases in the presence of the phase error.Inspired by the performance of the sparse autofocusing method to estimate both the scattering value and the phase error by an iterative optimization mean,the phase error estimation step in the sparse autofocusing method is implemented in the Fully Connected Layer and exploited into the LFISTA structure to propose the Tomo SAR imaging method termed as Ada TomoNet.Unlike the sparse autofocusing method,this method learns the phase error directly by network training without multiple iterations.Simulation experiments verify that the reconstruction accuracy of the proposed method can be improved by 10% compared with the LFISTA network-based imaging method and the conventional CS imaging method in the presence of phase error,and the reconstruction effect of the proposed method is found to be better than that of the LFISTA network-based imaging method and the CS imaging methods.2.A MB-Tomo SAR imaging method based on the FISTA-3DTV-Net network is studied.The conventional CS imaging model is based on the parametric canonical constraint,which mainly works on sparse strongly scattered targets and leads to the loss of weakly scattered targets during reconstruction.Considering that the space domain total variational(TV)norm has the property of maintaining target structure and edges and suppressing noise,this thesis establishes a joint L1-TV regularized laminar imaging model and solves it by the FISTA-TV algorithm;FISTA-TV is a fast solution algorithm for the joint L1-TV regularized model,which iterates by alternating variable optimization,firstly,the scene scattering coefficients are estimated,and the TV variables are estimated and then weighted to obtain the final estimated solution.However,this algorithm requires manual adjustment of the algorithm iteration parameters to obtain better results.To address this problem,this thesis proposes a layer analysis imaging method based on the FISTA-3DTV-Net network,which adds a TV variable update layer to the LFISTA The TV variable update layer is calculated in the same way as the TV variable in the FISTATV algorithm,with the difference that the network learns the optimal parameters through training,thus avoiding manually tuning of the parameters.Simulations and real SAR data experiments verify that the proposed method can recover the weak scattering target better and maintain the structural properties of the imaging results better than the FISTA-TV method and other CS methods.