Oceanic mesoscale eddy detection and convolutional neural network complexity

Abstract

Deep learning has drawn the attention of oceanographic researchers over the past few years, making the research community adopt computer vision techniques for oceanic mesoscale eddy detection on satellite altimetry gridded products. In this paper, we describe a convolutional neural network designed to detect eddies in satellite altimetry maps after being trained using segmentation masks provided by the OpenEddy detection algorithm. Against the current trend, in which increasingly complex neural networks are being proposed to address this problem, our design is relatively simple and yet provides competitive performance when compared to any of the previous deep learning methods reported in the literature. Furthermore, we show that our model is less sensitive to timely variations than the traditional models based on physical and geometric features defined by human experts, making it possible for our model to use the general data context to identify eddies that those traditional models would have missed. These results prove that overly complex neural network architectural designs are not required to solve the eddy detection problem on altimetry maps and generate a sufficiently good model for most practical applications in the field of marine sciences.