SAR speckle filtering with CNN using simulated training data

SAR sensors play an important role for a wide range of remote sensing applications due to their all-weather capabilities and their independence from sunlight. However, SAR images are more difficult to interpret than electro-optical images, and are affected by speckle noise, a multiplicative phenomenon that is created by the coherent nature of the image formation process and that might severely hinder image interpretation. Thus, speckle filtering is of great importance and many different approaches have been suggested. While most classical speckle filters use the local statistics in a sliding window, more recently non-local filtering and Convolutional Neural Networks (CNNs) have been used. The approach suggested in this paper falls into the latter category. For supervised CNN approaches, many training chips are necessary to teach the network the task it should solve. In the case of speckle filtering, many speckled/unspeckled image pairs are necessary. In this paper, we suggest creating training data for a U-Net architecture using SAR simulation. The simulation results of several large 3d scenes are shown, which in turn are used for the training of the U-Net for speckle filtering. First results on a TerraSAR-X image are shown and the potential and limitations of the current state of the approach are discussed.