Reconstruction of synthetic aperture radar data using hybrid compressive sensing and deep neural network algorithm

The reconstruction of reflectivity profile of the synthetic aperture radar (SAR) is an important field of research. SAR tomography is an advanced 3D imaging technique for the spectrum estimation in the elevation direction for each azimuth resolution cell. This work presents the processing chain for the tomographic reconstruction from ALOS PALSAR data for an urban region. First, the data are preprocessed by removing the speckle noise followed by atmospheric phase screen and topographic correction. Then the SAR images are stacked together with one master image and the remaining slave images on the baseline value. After the images are coregistered, the interferogram is generated from the image to obtain the difference of the phase value. Then the proposed super resolution SAR (SRS) algorithm is attempted for TomoSAR processing, which combines the functionality of modern machine learning method like deep learning with parametric block-based compressive sensing approach. Finally, a 3D image is reconstructed from the input data. Evaluation is carried out by comparing the results of the proposed method with other spectrum estimation methods such as nonlinear least square, Capon, and multisignal classification. The normal baseline of the interferometric fringes is about 368.54 m. The proposed SRS algorithm gives improved results with less mean elevation error of 1.8 m and the less standard deviation error of 4.85 m. Finally, the result reveals that the SRS algorithm performed better than other TomoSAR algorithms with the less relative error 0.003. In this work, the multitemporal SAR data are analyzed to estimate the height of the objects in the scene from the reconstructive reflectivity profile of the data. The results depict the efficiency of a hybrid algorithm that is both data driven and physics driven, improving the height estimation results with a relative error of 0.003.image