Advances in SAR speckle filtering using a CNN and simulated training data

SAR speckle filtering is an important pre-processing step for many downstream tasks, since all SAR images are affected by speckle due to the coherent imaging and image formation process. Several classical window-based speckle filters have been developed over the years, but recently convolutional neural networks (CNNs) have been applied to this problem. For the training of CNN-based speckle filters the question arises how training data should be generated, since for most approaches pairs of SAR images with and without speckle are needed. This paper describes the continued efforts to train a U-Net for speckle filtering using simulated SAR images with and without speckle as training data. For the simulations, the CohRaS® simulator of Fraunhofer IOSB has been modified to also create SAR-like images without speckle. The modification also includes a scheme to retain the mean gray value in homogeneous regions for the unspeckled versions of the simulations, since this is a desirable property of any speckle filter. The trained U-Net is then applied to the speckle filtering of a TerraSAR-X High Resolution Spotlight 300 (HS 300) scene and the properties of the filter are discussed in comparison with a classical Gamma-MAP filter.