A Feature Fusion Network for Multi-modal Mesoscale Eddy Detection

Abstract

As a marine phenomenon, mesoscale eddies have important impacts on global climate and ocean circulation. Many researchers have devoted themselves to the field of mesoscale eddy detection. In recent years, some methods based on deep learning for mesoscale eddy detection have been proposed. However, a major disadvantage of these methods is that only single-modal data are used. In this paper, we construct a multi-modal dataset containing sea surface height (SSH), sea surface temperature (SST) and velocity of flow(VoF), which are useful for mesoscale eddy detection. Moreover, we propose a feature fusion network named FusionNet, which consists of a downsampling stage and an upsampling stage. We take ResNet as backbone of the downsampling stage, and achieve multi-scale feature maps fusion via vertical connections. Additionally, dilated convolutions are applied in the FusionNet to aggregate multi-scale contextual information. Experimental results on the constructed multi-modal mesoscale eddy dataset demonstrated the superiority of FusionNet over previous deep models for mesoscale eddy detection.