Residual U-Net for SAR Image Despeckling – a Lightweight and Effective Deep Learning Approach
Abdullah Таher, Layan Al-Wadai, Riuof Al-Malki · 2025
Synthetic Aperture Radar (SAR) imaging offers high-resolution Earth observation capabilities regardless of weather or lighting conditions. However, the coherent nature of SAR signal processing significantly affects acquired images due to multiplicative speckle noise. This noise degrades image quality and hampers critical tasks such as target detection, classification, and scene understanding. This paper proposes an efficient deep learning-based despeckling approach that utilizes a residual U-Net architecture. The method aims to strike a balance between computational simplicity and effective noise suppression, making it suitable for practical SAR image analysis pipelines. By integrating residual learning into the U-Net structure, the model improves noise removal while preserving essential structural and textural details. The proposed framework is designed to be lightweight, robust, and adaptable to real-world SAR data without requiring extensive computational resources.