Ocean eddy automatic detection in satellite optical images of the Norwegian Sea using deep machine learning.

Ocean eddies are an important factor in the large-scale dynamics of the global ocean, including polar and subpolar regions. A robust statistical information on the number and characteristics of mesoscaleand submesoscale eddies will yield new insights on their effect on dynamics of large-scale currents, ice-edgevariability, and other dynamic and biochemical processes in the ocean. Optical images complement the results of the eddy identification study in radar and satellite altimetry images, each of which has its inherent limitations. In optical images, eddies are often observed as spiral or mushroom-shaped structures in chlorophyll distribution, which are formed through the effect of eddy rotation and convergence/divergence patterns. Massive studies of characteristics of ocean eddies require algorithms for their automatic identification. Although several such algorithms have been suggested for satellite altimetry and radar data, no such algorithm exists for satellite optical images. It this study we propose a machine deep learning algorithm for efficient automatic eddy detection in Sentinel-3 optical images. The Lofoten Basin of the Norwegian Sea, an area with a small Rossby deformation radius of less than 10 km, but densely populated with eddies, was selected as a region for algorithm training and validation. Even though the study area is known as one of the cloudiest areas of the northern polar latitudes, 52 mostly cloud-free images were collected over the 9 years of Sentinel-3 data, where 938 eddies were detected. For automatic eddy identification we used SegFormer neural network architecture with an AdamW optimizer, applied for 512×512 pixel tiles. In the course of validation high quality metrics were obtained: Precision = 0.94, Recall = 0.91, Intersection of Union = 0.87 and Dice = 0.93. This demonstrates high efficiency of the algorithm developed. The algorithm additionally identified several eddies missed during visual image inspection. The results of the study are particularly relevant to polar ocean regions, where the predominant eddy sizes are significantly smaller than in the tropics. The robust identification of eddies in optical images is a promising step forward in understanding mesoscale and sub-mesoscale eddy dynamics.