Content area
Full Text
1. Introduction
The South China Sea (SCS) is a large semienclosed sea where typhoons, mesoscale ocean eddies, internal waves, and other weather and marine phenomena occur frequently (Wang et al. 2014; Tuo et al. 2018). The population around the SCS is dense, and hence, the requirements for environmental and operational ocean forecasting are high. In 1992, ERS-1 and T/P remote sensing satellites observed near-global sea level anomaly (SLA) distributions through accurate continuous data, representing the first observing system that could be used for permit eddy-resolution global ocean forecasting (Smedstad et al. 2003). In recent years, various kinds of remote sensing, in situ observation and reanalysis datasets have been produced. However, SLA forecasting has not improved rapidly with the increase in data. SLA forecasting technology can be mainly divided into two types: numerical forecasting and empirical statistical forecasting. For the former, the most widely used global operational ocean models include the Hybrid Coordinate Ocean Model (HYCOM) (Chassignet et al. 2009) and the Nucleus for European Modeling of the Ocean (NEMO) (Madec et al. 2017). Currently, there is a Global National Real-Time Ocean Forecasting System (RTOFS) ocean model, which is based on eddy-resolving 1/12° global HYCOM and provides up to 8 days of forecasts using the daily initialization fields and a 3D multivariate data assimilation approach (Cummings 2006). The French Mercator Ocean International forecast systems use the NEMO to predict sea level values up to 10 days in advance (Drévillon et al. 2008). With the development of high-performance computing and observation systems, more scientific challenges have surfaced in terms of physical processes, parameterization schemes, and data assimilation algorithms for different numerical models (Bauer et al. 2015).
Deep learning has grown in popularity in recent years and has been applied to subgrid parameterization (Zanna and Bolton 2020), chaotic dynamical system forecasting (Pathak et al. 2018; Vlachas et al. 2018), XBT bias correction (Leahy et al. 2018), and ocean prediction (Zhang and Dai 2019; Song et al. 2020; Berbić et al. 2017). Recent studies have shown that both marine and meteorological forecasts are more accurate and energy efficient in terms of the parameterization of key physical processes (Jiang et al. 2018; Bolton and Zanna 2019; Gentine et al. 2018). Currently, both recurrent neural networks (RNNs) (Hochreiter and...