SAR Image Despeckling Using a Self-supervised Learning Model

Synthetic aperture radar (SAR) images are often contaminated by speckle noise. Reflectors, on the ground, return constructive and destructive interference within each SAR resolution cell can result in a multiplication noise known as speckle. Such noise affects subsequent computer data processing and human visual interpretation. To effectively reduce the impacts of noise on signals, recent deep learning-based methods have shown promising results in natural images. However, applying such methods directly in a SAR despeckling task is not feasible where no ground truth data are available for training. To address this issue, we adopt a self-supervised training strategy, where we only look at corrupted (noisy) data without explicit image priors or a likelihood model of corruption. A deep neural network is applied to learn a mapping to each pixel's noisy version from its neighborhood pixels. We use a random donut masking strategy on training samples to prevent the model from learning identity mappings, significantly improving training efficiency. Since the noise has independent characteristics, our model can recover the pixels’ clean version correctly without a clean reference during training. We quantitatively assess our proposed method's performance using natural images with ground truth information. Experimental results demonstrate improvements compared to conventional noise filters. Qualitative evaluation is also conducted using real-world SAR images, and our outcome shows better visual quality than those from conventional filters. Our despeckling self-supervised model suggests the potential of using deep learning for SAR image filtering in the absence of ground truth.