Detection of ocean eddies from satellite ocean color and SST measurements using a deep learning approach International Journal of Applied Earth Observation and Geoinformation

Mesoscale eddies can be captured not only by sea surface height (SSH) imageries derived from satellite altimetry missions, but also by ocean color and sea surface temperature (SST) imageries derived from the visible, near-infrared, shortwave infrared, and thermal bands of satellite sensors. Satellite SSH and SST data have been used for eddy detection in the past. With the development of the gap-free ocean color datasets, researchers have recently started to use satellite chlorophyll-a (Chl-a) data for eddy detection. In addition, satellite-derived diffuse attenuation coefficient at the wavelength of 490 nm [Kd(490)], which measures the water turbidity, is also found to be sensitive to the effect of mesoscale eddies. However, it has been found that a single parameter of SSH, SST or Chl-a is often not reliable to accurately detect ocean eddies. In this study, we propose a deep convolution neural network (CNN)-based model that incorporates three parameters, Chl-a, Kd(490), and SST, for eddy detection. In our implementation of neural networks, the image of each parameter of Chl-a, Kd(490), and SST runs through a separate encoding–decoding path to obtain its single-parameter based feature map of oceanic eddies. The three feature maps are then concatenated and convolved by a convolution layer, before the final “softmax” activation layer is applied to perform pixel-wise classification for eddy detection. The networks are tested in the Gulf of America (GoA) and the Gulf Stream Extension (GSE) region. It is found that the single-parameter of SST model does not always detect eddies in the GoA, while the single-parameter model of Chl-a or Kd(490) sometime misses eddy detections in the GSE region. However, three-parameter combined model is always reliable to accurately detect the anticyclone and cyclonic eddies of interests in the GoA and GSE. The eddy detection of the single-parameter models and the combined model are quantitatively evaluated, showing that the three-parameter combined model performed significantly better than all single-parameter models.