Data_Sheet_1.pdf

Abstract

Ocean mesoscale eddies constitute the ubiquitous, irregular, discrete components of water transport on a global scale. Accurately predicting the variation of ocean mesoscale eddies is of great scientific significance in understanding the evolution and propagation properties of mesoscale eddies. In this paper, we propose to make an initial attempt to explore spatio-temporal predictability of mesoscale eddies, employing deep learning architecture, which primarily establishes Memory In Memory (MIM) for sea level anomaly (SLA) prediction, combined with the existing mesoscale eddy detection. Oriented to the South China Sea (SCS) (125°-137.5°E, 15°-27.5°N), we quantitatively investigate the historic daily SLA variability at a 0.25° spatial resolution from 2000 to 2018, derived by multimission satellite altimetry, develop enhanced MIM prediction strategies on NVIDIA Titan Xp GPU, equipped with Gated Recurrent Unit (GRU) and spatial attention module, in a scheduled sampling manner. The gating mechanism of MIM-N and MIM-S has been improved to balance the model complexity and prediction accuracy, overcome gradient vanishing and explosion for long-term dependencies, and the spatial attention complements to strengthen feature extraction towards spatio-temporal variability in SLA, thus refraining from the connection structure mismatching within the convolution operation. At the early stage of training, the real value SLA input guides the model from initialization to a more reasonable state, while scheduled sampling gradually involve to intentionally feed the newly predicted value as the input, to tackle the distribution inconsistency between training and inference.It has been demonstrated in our experiment results that, from correlative steps, model construction, hyperparameter selection, pre-training, fine-tuning, optimization, transfer learning, performance evaluation, our proposed prediction scheme outperformed the state-of-art approaches for SLA time series, with MAPE, RMSE of the 14-day prediction duration respectively 5.1%, 0.023m on average, even up to 4.6%, 0.018m for the effective sub-regions, compared to 19.8%, 0.086m in ConvLSTM and 8.3%, 0.040m in original MIM, which greatly facilitated the mesoscale eddy prediction, in aid of classical mesoscale eddy detection.The proposed scheme could integrate evidences on the spatio-temporal responses of mesoscale eddies, help the observation deployment of ARGO, glider, AUV missions, and provide opportunities to potentially identify underlying patterns that mesoscale eddies involve in ocean dynamics.