TransSARNet: a deep learning framework for despeckling of SAR images

Abstract Synthetic Aperture Radar(SAR) images are extensively used for Earth observation because of their all-weather, day, and night imaging capabilities. However, speckle noise in SAR images significantly reduces their usability in a variety of applications. Deep learning models developed for SAR despeckling exhibit promising noise reduction capabilities. Bringing a balance between reducing graininess and preserving texture details is a challenging task. In addition, supervised training of a robust deep learning model requires noisy images that capture the SAR speckle dynamics and the corresponding speckle-free ground truth, which is generally not available. This study proposes the first hybrid CNN-Halo attention-based transformer model for SAR despeckling. CNN-based feature extraction modules provide multiscale and multidirectional and large-scale feature maps. A halo-attention transformer block is used in the skip connection. It aids in the better preservation of radiometric information in the despeckled SAR images. TransSARNet is trained in a supervised manner using a new synthetic SAR dataset, which is a combination of the Kylberg and UCMerced land-use datasets. This study also analyzed the effect of combining the Kylberg and UCMerced datasets on texture preservation in despeckled SAR images. The visual and qualitative metrics evaluated on Sentinel-1 Single Look Complex SAR data showed that the proposed TransSARNet approach outperformed the other models under consideration. TransSARNet achieves a harmonious balance between model complexity, despeckling ability, edge preservation, radiometric information preservation, and smoothing in homogeneous regions.