Transformer-Based Feature-Preserving Despeckling Framework for Synthetic Aperture Radar Imagery

Synthetic Aperture Radar (SAR) imagery is widely used in remote sensing because of its ability to acquire data under all weather and illumination conditions. However, SAR images are inherently affected by speckle noise, which reduces image clarity and makes the extraction of meaningful structural and textural information difficult. Effective despeckling methods must therefore suppress noise while preserving important image features such as edges and fine textures. In this work, a Transformer-Based Despeckling Network (TBDN) is proposed to address this challenge. The proposed architecture first extracts low-level features from the noisy SAR image using convolutional layers. These features are then processed through transformer encoder blocks that employ multi-head self-attention mechanisms to capture long-range spatial dependencies and contextual information. A multi-scale feature fusion module integrates hierarchical features to enhance structural preservation during noise removal. Finally, an image reconstruction layer generates the despeckled SAR image with improved visual quality. Experimental evaluation using standard metrics such as PSNR and SSIM demonstrates that the proposed model effectively reduces speckle noise while maintaining critical image details. The results indicate that transformer-based architectures provide a promising direction for advanced SAR image despeckling and feature-preserving image restoration.