Automatic Identification of Ocean Eddies in SAR Using U-Net Convolutional Neural Network

The effect of mesoscale and submesoscale eddies on regional heat and freshwater transport, sea-ice dynamics, and bioproductivity in polar and subpolar regions is largely unknown. The promising massive investigation of relatively small polar and subpolar eddies with Synthetic Aperture Radar (SAR) images requires robust automatic algorithms for their identification. Following recent advances in such algorithms for Sentinel-1 data, in this study describes an algorithm for identifying mesoscale and submesoscale eddies in earlier Envisat Wide-Swath SAR images using convolutional neural network based on U-Net architecture. The model was trained on 520 fragments of SAR images 512 × 512 pixels each, collected in the Greenland, Irminger and Labrador Seas. The resulting accuracy of eddy detection was 0.89. An improvement in the segmentation accuracy compared to previous studies was a result of the applied annotation methods, based on the analysis of typical texture patterns and morphological characteristics of ocean eddies in SAR. The developed algorithm showed its potential in identification of eddies not only in the Marginal Ice Zone, but also in the open water.