Deep Learning for Satellite Oceanography

We developed a W-Net architecture, a novel structure inspired from the classical U-Net architecture for semantic segmentation. Two U-net like branches are used, one for SST data and one for SSH data. The two branches give a segmentation with high accuracy for Gulf Stream and Rings. Using two parallel Encoder-Decoder networks (one branch for SST and the other for SSH), helped to obtain 82.7% raw test accuracy for Gulf Stream and a low error of 4.39% in the detected path length. For the Rings, W-Net obtained more than 71% raw eddy detection accuracy. No other model has been reported with such a high accuracy for this simultaneous estimation problem. This work is important for advancing automation in operational oceanography centres and is a big boost to use of AI and ML in ocean sciences.

Faculty: Deepak Subramani


D. Lambhate, R. Sharma, J. Clark, A. Gangopadhyay and D. Subramani, “W-Net: A Deep Network for Simultaneous Identification of Gulf Stream and Rings From Concurrent Satellite Images of Sea Surface Temperature and Height,” in IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-13, 2022, Art no. 4203213, doi: 10.1109/TGRS.2021.3096202.

Click image to view enlarged version

Scroll Up