As a longstanding scientific challenge, accurate and timely ocean forecasting has always been a sought-after goal for ocean scientists. However, traditional theory-driven numerical ocean prediction (NOP) suffers from various challenges, such as the indistinct representation of physical processes, inadequate application of observation assimilation, and inaccurate parameterization of models, which lead to difficulties in obtaining effective knowledge from massive observations, and enormous computational challenges. With the successful evolution of data-driven deep learning in various domains, it has been demonstrated to mine patterns and deep insights from the ever-increasing stream of oceanographic spatiotemporal data, which provides novel possibilities for revolution in ocean forecasting. Deep-learning-based ocean forecasting (DLOF) is anticipated to be a powerful complement to NOP. Nowadays, researchers attempt to introduce deep learning into ocean forecasting and have achieved significant progress that provides novel motivations for ocean science. This article provides a comprehensive review of the state-of-the-art DLOF research regarding model architectures, spatiotemporal multiscales, and interpretability while specifically demonstrating the feasibility of developing hybrid architectures that incorporate theory-driven and data-driven models. Moreover, we comprehensively evaluate DLOF from datasets, benchmarks, and cloud computing. Finally, the limitations of current research and future trends of DLOF are also discussed and prospected.

The El Niño–Southern Oscillation (ENSO) is the most dominant interannual variability of the ocean system exerting global environmental and socio-economic impacts through spatiotemporal teleconnections. Real-time ENSO prediction with physics-driven dynamical models suffers from large biases and uncertainties, while deep learning fails to consider time-varying spatial interactions across scales, causing difficulties in addressing the inherent complexity of teleconnections. Despite decades of efforts, predicting ENSO with lead times of more than one year remains challenging. In this study, we propose ENSOFarseer, an innovative deep learning framework that simulates cross-scale spatiotemporal teleconnections underlying ENSO dynamics and enhances long-lead predictability. ENSOFarseer outperforms state-of-the-art baselines, providing predictions at lead times of up to 24 months while mitigating the spring predictability barrier. ENSOFarseer accurately reproduces teleconnections by simulating interactions of local-scale feedback and global-scale teleconnections, offering insights into the contributions of various modes to predictions. The intrinsic improvement in long-lead prediction skill stems from the unified fusion of spatial multi-scale collaborative perception and temporal cross-scale dynamical evolution that tackles the inherent multiscale and multiphysics challenges and provides reliable probabilistic predictions by incorporating Bayesian uncertainty quantification. These findings highlight substantial improvements in cross-scale spatiotemporal deep learning in advancing ENSO prediction and deepening our understanding of its dynamics.

Sea surface temperature (SST) serves as a fundamental driver of ocean–atmosphere interactions and global climate variability, exhibiting strong nonstationarity, multiscale dynamics, and cross-variable coupling. However, current deep learning models often fail to capture these complex characteristics, limiting their ability to support accurate and physically consistent long-term SST prediction. To address these issues, we propose PAMSTnet, a unified deep learning framework for physics-informed adaptive multiscale spatiotemporal prediction. PAMSTnet leverages three-dimensional empirical wavelet transform (3DEWT) to learn interpretable multiscale spatiotemporal dynamics from raw observations, and applies multivariate spatiotemporal empirical orthogonal function (MSTEOF) to identify dominant cross-variable coupled modes. These physically meaningful representations are integrated into a deep ConvLSTM predictive network (DCPN) to support coordinated multiscale dynamical learning. Furthermore, PAMSTnet introduces physics-informed consistency learning (PICL) to enforce thermodynamic and dynamic constraints, enhancing physical consistency and interpretability. Extensive experiments demonstrate that PAMSTnet achieves superior performance against state-of-the-art baselines in long-term global SST prediction, reducing RMSE by 8.1% and improving ACC by 2.8% compared with the best-performing baseline, particularly under extreme climate events. Interpretation insights further highlight PAMSTnet’s adaptive representation of variable contributions and regional physical drivers. These findings position PAMSTnet as a promising paradigm for developing intelligent ocean prediction systems with enhanced physical consistency and interpretability.

Accurate reconstruction of ocean subsurface temperature (ST) is fundamental for understanding ocean dynamics and quantifying climate change. However, existing methods still face three major limitations: inadequate modelling of temporal multiscale variability, incomplete representation of spatial multiscale structure, and inefficient exploitation of coupled multivariate information. Here we propose a deep learning framework, STMSNet, that integrates multi-source remote sensing observations including sea surface temperature (SST), sea level anomaly (SLA), and sea surface wind (SSW) to reconstruct three-dimensional ST. To more effectively capture temporal multiscale variability, we first design a Temporal Scale Mixer (TS Mixer) based on temporal convolution, which enhances the representation of time-dependent features across distinct timescales. Second, to improve spatial multiscale feature learning and fusion, we develop a Spatial Feature Fusion Module (SFFM) built upon a Hierarchical Atrous Spatial Pyramid Pooling (HASPP) module. Third, to better exploit the coupled information among oceanic variables, we introduce an adaptive multi-task learning (MTL) architecture that uses salinity reconstruction as an auxiliary task, providing complementary information for the primary temperature reconstruction. This design enhances the model's accuracy and reliability. Experiments demonstrate that STMSNet reliably reproduces the spatial distribution and vertical structure of ST. The combination of temporal–spatial multiscale modules with the adaptive MTL framework markedly strengthens the representation of complex ocean variability. Relative to representative machine learning and deep learning baselines, STMSNet achieves consistently higher accuracy, particularly in dynamically active regions including internal wave regions, coastal upwelling zones, and mesoscale eddy fields. Overall, the proposed spatiotemporal multiscale reconstruction framework offers a promising route towards high-precision estimation of ocean ST.