Sea Ice SAR Image Generation and Analysis based on Generative Adversarial Networks

This work discusses the potential applications of deep learning method supported by the increasing amount of synthetic aperture radar (SAR) data, in particular the Generative Adversarial Networks (GANs). The generative and discriminative networks of a modified version named Wasserstein GANs (WGANs) are applied in the sea ice image analysis. The generated SAR images bear a great similarity to real images from both visual inspection and data modeling perspectives. Meanwhile, the discriminative network outputs are capable of providing direct representations of sea-ice regional proportions and indication of the overall ice concentration of the test images. Such results demonstrate a useful unsupervised or semi-supervised approach in sea characteristic studies. The network training and test scenes are from real RADARSAT -2 HH polarization SAR multi-looked intensity data covering marginal sea ice areas.