Sea and Land Segmentation Based on Pix2Pix Compared to FCM Using EOS-04 SAR Images

Sea/Land segmentation is of great significance for coastal research and maritime surveillance. Space borne Synthetic Aperture Radar (SAR) has the ability to penetrate clouds and thus can have high-resolution $24 \times 7$ imaging of large coastal areas under all weather conditions. However, SAR images of coastal areas usually have complex patterns because of the diverse topography and geomorphology, and the SAR imaging mechanism results in large coherent speckle noise in the images. In this study, we applied the Fuzzy C Means (FCM) and its variants as well as Pix2Pix GAN Model for sea-land segmentation and systematically investigated the models’ performances on Earth Observation Satellite (EOS04) images. FCM has been broadly used for semantic image segmentation as one typical approach of unsupervised learning. Compared to FCM, the GAN model can segment the sea and land more efficiently, and the results are very much encouraging and promising. Finally, we evaluated the accuracy of water-body detection using multiple performance indicators. The experiment demonstrates the high accuracy and reliability of our proposed sea-land segmentation method.