Graph-Based Memory Recall Recurrent Neural Network for Mid-Term Sea-Surface Height Anomaly Forecasting

Abstract

Sea surface height anomaly (SSHA) plays a pivotal role in ocean dynamics and climate systems. This article develops a graph-based memory recall recurrent neural network (GMR-Net) to achieve accurate and reliable mid-term spatiotemporal prediction of the SSHA field. The proposed method designs a newly developed long-term memory recall cell as the building block of the network, which utilizes the proposed memory store recall (MSR) module to learn and capture the mid- and long-term temporal dependencies of the SSHA field. The MSR module can efficiently recall memories stored in the memory bank across multiple timestamps through the proposed graph representation mechanism even after long periods of disturbance. The mid-term SSHA forecasting is performed with a 30-day ahead, and our proposed GMR-Net model achieves high prediction accuracy in different geographical regions: the Tropical Western Pacific and the South China Sea, yielding an RMSE of 0.026 and 0.035 m, respectively. Compared with advanced prediction models, our proposed GMR-Net model exhibits high reliability and superior performance in mid-term SSHA forecasting. Moreover, marine phenomena, such as Rossby waves, which can cause dramatic changes in sea-surface height, are successfully observed from our forecast data, further verifying the effectiveness of our prediction method.

Keywords

• Sea-surface height anomaly (SSHA)
• deep learning (DL)
• spatiotemporal prediction
• Rossby waves