Synthetic aperture radar imaging with deep neural networks

Official URL: https://risc01.sabanciuniv.edu/record=b2553782_(Table of contents)

Synthetic aperture radar (SAR) is a remote sensing imaging modality that has been in use since the 1960s. Conventional image formation in SAR is based on 2D inverse Fourier transform of the reflectivity field of the scene to be imaged. This conventional image formation technique is developed for a clean and complete data collection scenario. However, in reality, the collected data are only a reduced representation of the underlying scene due to hardware limitations and uncertainties in the data collection geometry, and hence suffer from reduction and phase errors. Therefore, many SAR image formation frameworks using regularization have been proposed over the years, in order to account for these limitations. In this dissertation, we have focused on the SAR imaging problem, particularly image formation, phase error correction, and automatic target recognition (ATR), and developed three frameworks. The first framework tackles the SAR image formation problem. In this framework, SAR image formation is formulated as a regularized optimization problem, and using the plug-and-play (PnP) priors framework, we have incorporated deep learning-based priors into our formulation. Our second framework is an extension of the first one, which aims at joint image formation and phase error correction. Experimental results show the effectiveness of these two frameworks and our proposed methods exceed the state-of-the-art image formation and phase error correction performances in the majority of the scenarios considered. The third proposed framework focuses on the ATR problem, and within this framework two ATR approaches are presented which perform the ATR task in the data domain rather than image domain. We have experimentally shown that the ATR task can be successfully performed in the data domain, and with further development, it might be possible to reach state-of-the-art performance. Overall, we have shown that the performance in various SAR imaging tasks can be improved significantly using deep learning tools.