Synthetic Aperture Radar information extraction and phase characterization via complex-valued neural networks
Date
2025
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Deep learning techniques based on optical imagery have demonstrated recent success with improving Synthetic Aperture Radar (SAR) image quality compared to traditional post-processing techniques. However, the capabilities and limitations of these neural network designs, specifically as applied to diverse SAR datasets, is lesser known. Furthermore, unlike optical imagery, SAR coherent collection also include phase information captured within the complex domain. This additional phase information is traditionally disregarded due to is random appearance and the need to fit real-valued neural network designs. However, this research demonstrates, that preserving and enhancing this phase information through single-channel complex-valued neural network can significantly improve SAR image enhancement, target characterization, and Moving Target Identification (MTI). ☐ To improve the dataset diversity and relevance for improved neural network training, a new high-resolution Sensor Independent Complex Dataset (SICD) was compiled and processed using Capella Space’s commercial SAR satellites operating in spotlight-mode [1]. Lower resolution training images were obtained through sub-aperture captures within the Fourier domain – creating a diverse and accurate training set with multiple lower resolution frequency and angular captures capable of being mapped to a single high-resolution complex scene. This sup-aperture sampling, along with pre-processing algorithms for reducing dataset noise, was shown to notably improve network performance. ☐ Initial amplitude domain research focused on comparing super-resolution deep learning architectures, based on Convolutional Neural Networks (CNNs), Residual Networks (ResNets), and Generative Adversarial Networks (GANs). These architectures were optimized to support SAR image enhancement using new performance metric evaluations, patch-wise statistics, 1D feature extractions, and detailed visual inspection. A real-valued Residual Regression CNN (RR-CNN) achieved greater speckle reduction, smoothing, and feature contrast, with residual blocks offering improvements for network expansion. In comparison, a real-valued Conditional Cycle GAN (CC-GAN), with the key addition of L1 loss and cycle-consistency loss, significantly improved scattering point separations and reduced ringing artifacts. This research provides direct comparisons for neural network designs, loss-functions, and hyperparameter selections, and establishes amplitude-domain deep learning recommendations for super-resolving SAR imagery. While promising image enhancement improvements for SAR images were achieved through these amplitude-domain deep learning techniques, these studies were further extended into the complex-domain to fully exploit the additional phase information available with SAR coherent collections. ☐ This phase information was found to provide important insight on scattering behavior that can further improve SAR deep learning analysis and automatic feature identification, as shown through phase derivative calculations. Novel single-channel complex-valued neural networks (CVNNs) were designed and optimized to preserve and enhance SAR phase characterization. These evaluations focused on extending the amplitude-domain CNNs, ResNets, and GANs into the complex-domain to support complex-valued regression outputs, convolutions, weights, normalization layers, and activation functions. Loss functions for complex-valued inputs were also evaluated for phase preservation. The optimized CVNNs were directly compared to the similar Real-Valued Neural Networks (RVNNs), with the CVNNs achieving greater speckle reduction and higher peak signal strengths. 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 was further demonstrated through improved phase derivative separations, in which the target feature phase sensitivity even surpassed the original high-resolution captures. These learnings provide a new complex-valued approach for super-resolving SAR imagery and improving target phase characterization. ☐ To further explore this, CVNNs were compared to RVNNs for moving target characterization and identification within high clutter scenes. SAR coherent collections traditionally focus on stationary artifacts, resulting in the distortion of non-stationary processes and limiting the identification of moving targets within background clutter. To provide automated MTI for traditional single-channel SAR collection modes, a new CVNN was developed that improves detections within high clutter environments. To support this analysis, a new labeled complex-valued dataset was created using synthetically injected moving targets within the Compensated Phase History Data, allowing for various target-to-clutter signatures, target headings, and target speeds. Additionally, CVNN comparisons were provided for both the complex spatial and Fourier domains. Phase derivative analysis visualized the enhanced moving target characterizations achievable within the complex domain, while CVNN MTI analysis quantified these results, achieving a target detection accuracy of 81.4%, compared to 77.0% for a similar RVNN. Furthermore, for the lowermost 10% of target-to-clutter signatures, the CVNN correctly classified 57.7% more targets compared to the RVNN. The CVNN also detected 10.1% more targets with speeds exceeding 20 knots and 11.1% more targets with predominately cross-range velocities. These results highlight the ability for CVNNs to also improve MTI sensitivities beyond those achievable in only the amplitude domain. The results from this research establish neural network processing recommendations, as extended into the complex-domain, for improving the information extraction from SAR collections.
Description
Keywords
Complex-valued neural networks, Dataset generation, Image enhancement, Moving target indication, Phase characterization, Synthetic Aperture Radar