Deep Learning for Detecting Gulf Stream and Eddies from Satellite Images

Deepak Subramani, Indian Institute of Science, Computational and Data Sciences, Bangalore, India, Raghav Sharma, Indian Institute of Science, Bangalore, India and Avijit Gangopadhyay, Professor of Oceanography, School for Marine Science and Technology University of Massachusetts Dartmouth 836, S. Rodney French Blvd. New Bedford, MA, New Bedford, MA, United States
Abstract:
Accurate digitization of synoptic ocean features is crucial for operational ocean forecasting needs. For example, these digitized features are utilized for data-assimilation and in feature oriented regional modeling systems. Currently, skilled human operators visualize and extract these features through a time-consuming manual process. Our motivation is to automate this task and develop an automated data-driven feature oriented regional modeling and data assimilation system. Towards this end, we develop a deep learning method to automatically detect and extract the position of the Gulf Stream, and warm and cold eddies from satellite images of sea surface temperature and sea level anomaly. First, a labelled data-set of the position of Gulf Stream and eddies are prepared from the weekly images of Jennifer Clark’s Gulf Stream for the years 2014 to 2017. The corresponding SST map is obtained from the freely available AVHRR imagery processed by John Hopkin’s University Applied Physics Laboratory. A deep convolutional neural network is trained with SST as input and the GS and eddy positions as output. We perform SST data augmentation and where available Sea Level Anomaly data is provided for information augmentation. The data-set is carefully split to test and train so that both contain representative seasonal variations. The final U-net can perform this three-class segmentation task to high degrees of accuracy both on the test set and on the validation set of images (2018). We showcase and analyze results from applying our network to detect GS and eddy positions on un-labelled 2019 SST data.