Exploiting Aerial Imagery for Supervised Learning of SAR Despeckling Neural Networks

Many applications utilizing SAR data, such as change detection, segmentation and classification, are impaired by the multiplicative speckle interference inherent in the imagery. Thus despeckling of SAR imagery is a often the key to developing robust algorithms for scene understanding. In recent years numerous deep learning-based approaches to speckle reduction have been proposed. However, the performance of these methods has largely failed to meet the expectations of researchers and industry alike. A key reason for this is due to the lack of accurate SAR-based ground truth training data. In this paper we propose the use of very high-resolution (VHR), low speckle aerial imagery and an accurate speckle model, as a ground truth signal for training a despeckling network based on the DnCNN architecture. Furthermore, we investigate modifications to the training formulation and finally demostrate the approach on Capella-2 VHR X-band imagery.