Synthetic aperture radar image enhancement and phase characterization using complex-valued neural networks

Deep learning image enhancement improvements for synthetic aperture radar (SAR) have primarily treated SAR imagery similar to optical imagery, disregarding the additional phase information available within SAR coherent collections. Through phase derivative calculations, we show this phase information to provide important insight into scattering behavior that can further improve SAR image quality and target characterization. Single-channel complex-valued neural networks (CVNNs) were designed and optimized to enhance SAR phase information. These evaluations focused on convolutional neural network, residual network, and generative adversarial network designs that fully support complex-valued regression outputs, convolutions, weights, normalization layers, activation functions, and loss function inputs. The CVNNs were directly compared with similar real-valued neural networks, with the CVNNs achieving greater speckle reduction and lower noise floors. This resulted from a reduction in phase interference effects as the CVNNs learned the structured phase response shared between the low-resolution and high-resolution captures, effectively reducing the random background interference. This also resulted in improved phase derivative separations for key features that even surpass the high-resolution captures. These learnings provide a new complex-valued approach for super-resolving SAR imagery and improving target phase characterization, with findings that can also extend into automatic target recognition algorithms.