PolSAR Image Classification With Complex-Valued Diffusion Model as Representation Learners

Polarimetric synthetic aperture radar (PolSAR) image classification is a momentous task in remote sensing domain. Recently, the explosive development of deep learning (DL) has dramatically boosted the performance of PolSAR image classification. However, existing DL-based methods are confronted with two barriers: 1) It is difficult to obtain sufficient labels that supervised DL models depend on; 2) The complex-valued characteristic of PolSAR data, especially the phase information, was underexplored in prior research. To tackle these two issues, we propose a generative self-supervised PolSAR image classification approach based on a complex-valued diffusion model. Specifically, a diffusion-based polarimetric representation learning framework is designed, which is comprised of a forward noise diffusion process and a reverse denoising process. Via progressively adding noise to clean PolSAR data and recovering the original input with deep neural networks, the latent noise distribution and further discriminative polarimetric features can be extracted without any manual labels. The learned diffusion features are then fed into a classification module for land cover classification. It should be highlighted that the noise distribution and networks are all designed in complex domain, harnessing both the amplitude and phase information of PolSAR data. Experiments on four benchmark PolSAR datasets reveal the strong representation learning capability of diffusion models in PolSAR image classification and proffer insights to the values of generative models beyond conventional data synthesis in remote sensing realm.