AI Contributes to Ocean Eddy Detection by Synthetic Aperture Radar
May 20, 2024
Scientists at the Aerospace Information Research Institute (AIR) under the Chinese Academy of Sciences (CAS), in collaboration with their international collogues, have proposed a state-of-the-art deep-learning network, named EOLO, aiming to improve the detection of ocean eddies observed in C-band spaceborne synthetic aperture radar (SAR) imagery. This work has been published in Remote Sensing of Environment, which combines advanced AI algorithms with high-resolution spaceborne SAR data, providing a methodological basis for further studies of sub-mesoscale eddies.
Ocean eddies are rotary currents of water, which play a crucial role in the global energy cycle and significantly influence the transport of heat, salt, and nutrients across vast marine regions. Eddies with diameters larger than the first baroclinic Rossby radius are known as mesoscale eddies, and those with diameters smaller than this radius are called sub-mesoscale eddies. During the past decades, radar altimeters (RAs) have become powerful enough to help scientists investigate mesoscale eddies, while sub-mesoscale ones have been challenging due to their small size and short lifetimes.
SAR, with its high spatial resolution and its independence of daytime and weather conditions, has shown its unique capability in sub-mesoscale eddy observation. This study developed an algorithm for automatic detection of ocean eddies observed in SAR imagery and extraction of their spatial features, called EOLO (Eddy detection based on the YOLO algorithm).
Based on the ocean eddy dataset (named EddyDataset) established from Sentinel-1 (S1) SAR data, a series of improvements, e.g., channel attention mechanism, new feature fusion method, etc., were introduced and finally obtained a high-quality eddy detection network with 91.5% average precision. To further validate the generalization of the EOLO, it was applied to a large number of randomly selected SAR data acquired over the Red, Baltic, and Western Mediterranean Seas, achieving 96.6%, 98.8% and 98.9% precision, respectively. This suggests that EOLO has good generalization ability.
Moreover, based on the S1 data acquired over the Western Mediterranean Sea in 2021, 8056 SAR eddies were detected by the EOLO network. Comparison with the eddies identified by RA in the same period shows that there is a remarkable difference in the scale of the ocean eddies observed by SAR and RA. Typical diameters of SAR eddies range from 2 km to 20 km while those of RA eddies range from 50 km to 170 km. The quantitative analysis found that about 46% of the ocean eddies, mainly sub-mesoscale eddies, cannot be observed by existing radar altimeters. This provides important observational facts for further research on the dynamics of sub-mesoscale and mesoscale ocean eddies.
With its high precision and efficiency, EOLO is expected to advance our understanding of ocean dynamics, climate patterns, and marine ecosystems, ultimately contributing to more accurate environmental predictions and informed decision-making for sustainable ocean management.
“In the next step, the team will continue to improve the generalization capability of the model, and then realize the automatic detection of ocean eddies on multi-band spaceborne SAR imagery. On this basis, the team will try to break through the difficulty of the three-dimensional perspective of sub-mesoscale eddies”, said Prof. LI Xiaoming, the correspondence author of the study.
Research News
AI Contributes to Ocean Eddy Detection by Synthetic Aperture Radar