Abstract
Automatic detection of mesoscale oceanic eddies is in great demand to monitor their dynamics which play a significant role in ocean current circulation and marine climate change. Traditional methods of eddies detection using remotely sensed data are usually based on physical parameters, geometrics, handcrafted features or expert knowledge, they face a great challenge in accuracy and efficiency due to the high variability of oceanic eddies and our limited understanding of their physical process, especially for rich and large remotely sensed data. In this paper, we propose a simple deep architecture DeepEddy to detect oceanic eddies automatically and be free of expert knowledge. DeepEddy can learn high-level and invariant features of oceanic eddies hierarchically. It is designed with two principal component analysis (PCA) convolutional layers for eddies feature learning, a binary hashing layer for non-linear transformation, a feature pooling layer using block-wise histograms and spatial pyramid pooling to resolve the complicated structures and poses of oceanic eddies, and a classifier for the final eddies identification. We verify the accuracy of the architecture with comprehensive experiments on high spatial resolution Synthetic Aperture Radar (SAR) images. We achieve the state-of-the-art accuracy of 96.68%.