Wrapped Phase Denoising Using Denoising Diffusion Probabilistic Models

Wrapped phase denoising is a crucial step in the interferometric synthetic aperture radar (InSAR) data processing workflow. Traditional heuristic algorithms overly rely on designers’ experience and struggle to cope with complex and variable noise environments. In contrast, the convolutional neural network (CNN) and generative adversarial network (GAN) methods widely used in recent years have brought about challenges such as excessive smoothing or difficulties in training. This letter introduces a denoising method for the wrapped phase based on the diffusion model. The approach comprises a forward and reverse process, with its essence lying in training a noise removal network. This network gradually restores random noise to its corresponding clean phase using the noise phase as a conditional input. Furthermore, a novel conditional input scheme based on the sine and cosine of the phase enhances the training and sampling processes. The experimental results demonstrate that the method presented in this letter outperforms the other three methods in both simulated and real scenarios. Specifically, in the denoising of real interferometric phases (size $512\times 512$ ), the residue removal rate reaches 99.71%, showcasing remarkable denoising performance.