Due to a unique imaging mechanism, Synthetic Aperture Radar (SAR) images typically exhibit degradation phenomena. To enhance image quality and support real-time on-board processing capabilities, we propose a lightweight deep generative network framework, namely, the Lightweight Super-Resolution Generative Adversarial Network (LSRGAN). This method introduces Depthwise Separable Convolution (DSConv) in residual blocks to compress the original Generative Adversarial Network (GAN) and uses the SeLU activation function to construct a lightweight residual module (LRM) suitable for SAR image characteristics. Furthermore, we combine the LRM with an optimized Coordinated Attention (CA) module, enhancing the lightweight network’s capability to learn feature representations. Experimental results on spaceborne SAR images demonstrate that compared to other deep generative networks focused on SAR image super-resolution reconstruction, LSRGAN achieves compression ratios of $74.68\%$ in model storage requirements and $55.93\%$ in computational resource demands. In this work, we significantly reduce the model complexity, improve the quality of spaceborne SAR images, and validate the effectiveness of the SAR image super-resolution algorithm as well as the feasibility of real-time on-board processing technology.

Due to its capability for all-weather, all-time information acquisition, Synthetic aperture radar (SAR) plays a vital role in the field of Earth observation. However, the specificity of the radar sensor and the complexity of electromagnetic scattering imaging physics result in SAR images lacking the intuitiveness of optical images, making them unsuitable for interpretation by non-experts. A common approach to tackle this challenge is to use conditional generative adversarial network (cGAN) to translate SAR images into optical images, thereby enhancing readability and assisting non-experts in interpretation while filling the gaps in optical data due to acquisition constraints. Nevertheless, traditional cGAN-based methods are limited by inadequate global semantic information extraction and poor detail preservation, leading to translated images with incoherent texture and color, and blurred edge. To address these issues, we propose a hybrid vision transformer conditional generative adversarial network (HVT-cGAN) for SAR-to-optical image translation (S2OIT). In our proposed HVT-cGAN, the generator utilizes a convolutional stem for patch embedding and encoding. The parallel CNN branch and vision transformer branch are employed for the extraction and mapping of local and global information, respectively. Moreover, we propose a novel attention-based feature fusion module, named the Convolutional Attention Fusion Module (CAFM), which can adaptively aggregate local and global information from parallel branches by learning both channel-wise and spatial-wise relations. Benefiting from these improvements, our method achieves superior performance in both qualitative and quantitative comparisons with other methods on the SEN1-2 dataset. Additionally, the results of multiple ablation experiments validate the effectiveness of the proposed method.

Compared to traditional real-valued neural networks, which process only amplitude information, complex-valued neural networks handle both amplitude and phase information, leading to superior performance in polarimetric synthetic aperture radar (PolSAR) image classification tasks. This paper proposes a multi-scale feature extraction (MSFE) method based on a 3D complex-valued network to improve classification accuracy by fully leveraging multi-scale features, including phase information. We first designed a complex-valued three-dimensional network framework combining complex-valued 3D convolution (CV-3DConv) with complex-valued squeeze-and-excitation (CV-SE) modules. This framework is capable of simultaneously capturing spatial and polarimetric features, including both amplitude and phase information, from PolSAR images. Furthermore, to address robustness degradation from limited labeled samples, we introduced a multi-scale learning strategy that jointly models global and local features. Specifically, global features extract overall semantic information, while local features help the network capture region-specific semantics. This strategy enhances information utilization by integrating multi-scale receptive fields, complementing feature advantages. Extensive experiments on four benchmark datasets demonstrated that the proposed method outperforms various comparison methods, maintaining high classification accuracy across different sampling rates, thus validating its effectiveness and robustness.

Compressed sensing-based synthetic aperture radar (SAR) sparse imaging technique has been widely used in SAR reconstruction tasks, but it is limited by high computational costs and poor imaging performance. In this letter, we propose a dynamic deep unfolding network based on content-aware shrinkage for SAR sparse imaging, called CAS-D²UN. First, an adaptive soft shrinkage mechanism is introduced to obtain content-aware threshold and enhance the adaptability of the model. Second, a dynamic path-controllable solution is proposed to adjust the network structure based on the complexity of the input data, thereby improving inference speed. Finally, experimental results demonstrate that our proposed CAS-D²UN can achieve superior sparse reconstruction for complex scenes while shortening the imaging time.

Synthetic Aperture Radar (SAR) imagery is inherently corrupted by multiplicative speckle, which degrades visual quality and impedes downstream interpretation. While deep learning has shown promise, supervised approaches require large-scale paired noisy-clean datasets that are difficult to acquire, and self-supervised schemes often rely on simplistic noise models incompatible with SAR physics. To overcome these challenges, we introduce DeCo-Despeckle, a novel and interpretable unsupervised framework that learns to despeckle using only unpaired SAR images. The method is grounded in the explicit decoupling of an image into its content and speckle components via a prior-guided, dual-branch architecture. This framework is optimized through a synergistic combination of learning objectives: an adversarial loss drives the output towards perceptual realism, while reconstruction and statistical priors enforce physical consistency. Crucially, a powerful multilayer patch-wise contrastive loss is integrated to ensure the model produces structurally accurate results by preserving high-fidelity details. Extensive experiments on real-world datasets demonstrate that DeCo-Despeckle not only establishes a new state-of-the-art for unsupervised despeckling but also outperforms several fully supervised methods in both quantitative metrics and visual quality, highlighting its practical utility and robustness.