A neural network approach for remote detection of marine eddies

This paper presents a machine learning approach for detection of Mediterranean water eddies from sea surface temperature maps of the Atlantic Ocean. Two methods based on texture analysis of the satellite imagery are evaluated. Given a map point, the first method extracts information on the surrounding thermal gradient and arranges it as a numerical vector of gradient angles. The second method uses Laws' algorithm to create a vector of numerical measures of structural features. In both the cases, a neural network is trained to recognise those numerical patterns that reveal the presence of eddy structures. Both the algorithms achieve high recognition accuracy and fast and robust learning results. Particularly important are the very low rates of false detections obtained, since eddies occupy only a small portion of the ocean area. Compared to Laws' method, the gradient-based algorithm gives comparable recognition accuracies with a lower design effort and at reduced computational costs. The simple and modular structure of the gradient-based method also compares favorably to the complexity other algorithms for identification of marine phenomena published in the literature. Given the competitive accuracy results obtained, the gradient-based approach may be preferable to the currently employed techniques since it is simpler and more easily reconfigurable.