Self-supervised global − local collaborative network for real SAR despeckling

Synthetic Aperture Radar (SAR) is crucial for Earth observation because it can acquire high-resolution images in all weather conditions. However, the presence of speckles—an inherent multiplicative noise caused by the coherent imaging process—severely degrades image quality and impairs the performance of subsequent interpretation tasks. To effectively capture both global contextual cues and fine-grained structural details in SAR image despeckling, we design a dual-branch Global-Local Collaborative Network (GLCNet) based on blind-spot convolution. GLCNet is trained in a self-supervised manner, requiring only original images for learning, making it well-suited for SAR data without ground truth. In the global branch, the SAR image is first decomposed into multiple frequency sub-bands through a Wavelet-Shuffle Downsampling (WSD), which decorrelates speckle components across scales and frequencies. A multi-scale blind-spot convolution is then applied to each sub-band in parallel, enabling the extraction of global textures without introducing speckle bias. In contrast, the local branch focuses on structure-aware restoration by jointly modeling frequency and spatial priors. By leveraging neighboring-pixel dependencies, this branch enhances local detail recovery and edge sharpness. Finally, an adaptive Detail-Guided Module (DGM) dynamically integrates complementary features from both branches, ensuring a harmonious balance between texture smoothness and structural fidelity. The proposed method is validated using various SAR sensors, including Sentinel-1, GF-3, TerraSAR-X, and Capella-X, demonstrating its superiority over traditional and deep learning approaches. Additionally, the application analysis confirms that the method enhances both the visual quality and analytical reliability of SAR images, making it a valuable preprocessing step for real-world scenarios. For reproducibility, our code and data are available at https://github.com/yangyang12318/LGCN.