Short- to Medium-Term Sea Surface Height Prediction in the Bohai Sea Using an Optimized Simple Recurrent Unit Deep Network

Abstract

Global warming has intensified the rise in sea levels and has caused severe ecological disasters in shallow coastal waters such as the Northeastern China's Bohai Sea. The prediction of the sea surface height anomaly (SSHA) has great significance in the context of monitoring changes in sea levels. However, the non-linearity of SSHA due to the occurrence of dynamic physical phenomena poses a challenge to current methods(e.g., ROMS, MITgcm) that aim to provide accurate predictions of SSHA. In this study, we have developed an optimized Simple Recurrent Unit (SRU) deep network for the short- to medium-term prediction of the SSHA using Archiving Validation and International of Satellites Oceanographic (AVISO) data. Thanks to the parallel structure of the SRU, the computational complexity of the deep network can be reduced to a considerable extent and this makes the short- to medium-term prediction more efficient. To avoid over-fitting and a vanishing gradient, a skip-connection strategy has been utilized for model optimization, and this improves significantly the accuracy of prediction. Detailed experiments were carried out in the Bohai Sea to evaluate the proposed model and it was demonstrated that the proposed framework (i) outperformed significantly the current deep learning methods such as the BP (Backpropagation), the RNN (Recurrent Neural Network), the LSTM (Long Short-term Memory), and the GRU (Gated Recurrent Unit) algorithms for 1, 5, 20, and 300-day prediction; (ii) can predict the short-term trend in the SSHA (for the next day or 2 days) in real time; and (iii) achieves medium-term prediction in seconds for the next 5–20 days and shows great potential for applications requiring medium- to long-term predictions. To the best of our knowledge, this is the first paper that investigates the effectiveness of the SRU deep learning model for short- to medium-term SSHA predictions.