Robust SAR Image Despeckling by Deep Learning From Near-Real Datasets

The inherent speckle in synthetic aperture radar (SAR) images significantly affects their potential usefulness, and its effective suppression is a challenging and nontrivial task. This article uses near-real SAR intensity datasets as the training data for the first time and proposes a robust deep learning-based speckle removal model: phase-guided deep despeckling network (PGD2Net). Owing to the unique geometric distortions as well as complicated speckle noise in SAR images it is extremely difficult to simulate geometric distortion similar to real SAR image. In addition, simulating noise as uniform and single-distributed also fails to fully represent speckle complexity. This article uses the temporal and spatial information of time series SAR images to create near-real SAR intensity datasets using an adaptive multilook method called generalized likelihood ratio test, which outstandingly solves the problems encountered with simulated data. Based on the correlation between intensity and phase, to improve the accuracy of speckle noise estimation, we introduce phase information in one subnetwork (speckle noise estimation subnetwork) of the proposed PGD2Net. Our ablation experiments demonstrate that this further enhances the network's speckle suppression performance. Moreover, we establish another subnetwork (dual-branch denoising subnetwork) to conduct feature interaction and estimate the clean intensity image based on a specially designed cross-attention module. Quantitative and qualitative results demonstrate that, compared to other algorithms, our proposed method exhibits strong adaptability and performance across scenes with varying degrees of geometric distortion and speckle due to different terrain undulations. Simultaneously, we extend this algorithm to SAR data obtained from different sensors, achieving excellent execution performance as well.