Artificial intelligence in oceanography has demonstrated a great potential with the explosive growth of ocean observation data and numerical model products. This article first reviews the history of ocean big data development, and then introduces in detail the current status of artificial intelligence in oceanography applications including identifying ocean phenomenon, forecasting ocean variables and phenomenon, estimating dynamic parameters, correcting forecast errors, and solving dynamic equations. Specifically, this article elaborates the research on the intelligent identification of ocean eddies, internal waves and sea ice, the intelligent prediction of sea surface temperatures, El Ni?o-Southern Oscillation, storm surges, waves and currents, the intelligent estimation of ocean turbulence parameterization for numerical models, and the intelligent correction of waves and current forecast errors. In addition, it discusses the recent progress of applying physical mechanism fusion and Fourier neural operator for solving ocean dynamic equations. This article is based on the current status of artificial intelligence in oceanography and aims to provide a comprehensive demonstration of the advantages and potential of applying artificial intelligence methods in the field of oceanography. With the two emerging research hotspots: digital twin oceans and artificial intelligence large models, the future development direction of artificial intelligence provides enlightenment and reference for interested scientists and researchers.

Waves are one of the most important phenomena in the ocean. The accurate and quick updated wave forecasting is of crucial significance for ensuring marine activities safety. The development of wave forecast is presented, including the traditional statistical wave forecasting methods, numerical wave prediction models, and the rapidly developing artificial intelligence (AI) wave forecasting methods. Currently, AI wave forecast models have been demonstrated unique advantages in terms of computational efficiency and adaptive forecasting accuracy, and they are gradually being applied in practical wave forecasting operations, transitioning from the research stage. However, they also have limitations, including limited forecasting elements, underestimation of extreme wave conditions, and weak forecasting generalization ability. Based on the characteristics of AI wave prediction, key scientific and technological issues that need to be addressed in current AI wave forecasting are proposed. These include efficient utilization of observational data, incorporation of prior physical knowledge, and enhancement of AI model safety and generalization ability.

Ocean eddies are prevalent oceanic phenomenon that play a crucial role in the global transportation of oceanic materials and energy. Although traditional methods for detecting ocean eddies are widely used, they suffer from significant drawbacks such as excessive reliance on expert-set thresholds, continuous manual intervention, large detection errors, low efficiency, and poor global applicability, making it difficult to adapt to the complex and variable marine environment. Currently, the rapid development of artificial intelligence (AI) presents a promising solution for the intelligent detection of ocean eddies. AI can automatically and rapidly extract deep features from images, effectively address the challenges posed by the high similarity in oceanic phenomenon features and significant geometric variability. This paper provides an overview of AI-based oceanic eddy identification methods based on different deep learning methods, focuses on coder-decoder structure, fully convolutional neural network, multi-scale context method and attention mechanism, and aims to provide valuable insights and references for future ocean eddy research.

The Indian Ocean Dipole (IOD) is a pivotal climate phenomenon in the Indian Ocean region, exerting a significant impact on the climate change of the surrounding areas and the global climate system. Accurate prediction of IOD is essential for comprehending the dynamics of the global climate, yet traditional forecasting methods are limited in capturing its complexity and nonlinearity, constraining predictive capabilities. This paper begins by outlining the relevant theories of IOD and evaluates the strengths and weaknesses of traditional forecasting methods. It then provides a comprehensive analysis of the application and development of deep learning in the field of IOD prediction, highlights the advantages of deep learning models over traditional methods in terms of automatic feature extraction, nonlinear relationship modeling, and large data processing capabilities. Additionally, the paper discusses the challenges faced by deep learning models in IOD forecasting: including data scarcity, overfitting, and model interpretability issues, and proposes future research directions to promote innovation and progress in the application of deep learning technology in the field of climate prediction.

Histopathological indexes can be used to assess health of organism, but their application faces challenges such as low efficiency, high cost and strong subjectivity. Introducing artificial intelligence (AI) technology into histopathological analysis of biological tissues can leverage its high-throughput image analysis capabilities, overcoming the limitations in assessing and monitoring marine organism health. Based on our review on health assessment indicators of marine organisms, the application of AI in image analysis, and the use of AI for histopathological image processing, a deep learning-based histopathological image analysis approach was proposed using gill tissues of marine mussels as representative. Through a series of processes such as training, validation, and prediction of histopathology images, it was determined that the Res-UNet deep learning model can efficiently and accurately quantify histopathological damage in mussels’ gills. An automated, high-throughput, and less subjective workflow based on deep learning was finally established, offering new ideas and techniques for marine organism health assessment and marine monitoring.

An accuracy analysis of wind speed data retrieved from Sentinel-1 synthetic aperture radar (SAR) was conducted based on buoy observations from the National Data Buoy Center (NDBC). A back propagation (BP) neural network was utilized to correct the deviation in the SAR-derived wind speeds. Sensitivity experiments were designed for environmental factors, the number of training samples for BP neural network input, and neural network structure parameters. Finally, the SAR wind field data were converted into u and v vector wind data, and the accuracy analysis and correction were performed separately for u and v wind components. The experiment finds that the SAR-derived wind speed is underestimated compared to the buoy data. After calibration using BP neural network, the accuracy of SAR-derived wind speed data is improved, and the absolute value of bias of wind speed decreases from 0.78 m/s to 0.04 m/s, the RMSE of wind speed decreases from 1.98 m/s to 1.77 m/s. The sensitivity experiments suggest that low quality environmental factors input data will decrease the calibration effect of BP neural network, and increasing the sample size of the training set can improve that. The calibration results of converted u and v vector wind field data also show that the BP neural network has good correction effect.

Compared with the semi-theoretical and semi-analytical wave forecasting and numerical modeling,artificial intelligence wave forecasting has the advantages of higher forecasting accuracy and lower computational resource requirements. In this paper, a two-dimensional significant wave height (SWH) forecasting model for the southeast coast of China is established based on the convolutional long short-term memory network (ConvLSTM) algorithm using ERA5 (ECMWF Reanalysis v5) reanalysis data as the initial field. The data from 2014-2022 are used to train the forecasts of SWH for the next 6 h, 12 h, 18 h and 24 h, and the data from 2023 are used for testing. Sensitivity tests are carried out to optimize the model configuration and evaluate the prediction performance of SWH in the southeast coast of China at four period validity (6 h, 12 h, 18 h, 24 h) in 2023. Sensitivity tests show that when input time series length N=4(input SWH value of -18 h, -12 h, -6 h, 0 h), the accuracies of the model at four period validity are better than those of other time series length. When the combination of input physical elements is SWH, mean wave direction and sea surface 10 m wind vector, the accuracy of the model is better than other combinations at 12 h, 18 h and 24 h. Through the fine-tuning of ConvLSTM model training and configuration, the two-dimensional and high-precision intelligent prediction of SWH in the southeast coast of China can be realized.

Extremely deep convective clouds are a precursor to the rapid intensification (RI) of tropical cyclones (TCs). To predict the RI of TCs in the western North Pacific (WNP), a machine learning (ML) model using data related to extremely deep convective clouds was developed. The ML model integrates the Statistical Hurricane Intensity Prediction Scheme-Rapid Intensification Index (SHIPS-RII) data and the coverage area of extremely deep convective clouds within a 300 km radius of the center of the TC. Based on data from 2011 to 2019, the model forecasted RI events that increased by more than 30 kn and 35 kn within 24 h. Compared to models using only SHIPS-RII data, this ML model showed an improvement of 5.66% and 9.58% in the Peirce Skill Score (PSS), and a relative increase of 8.41% and 8.55% in the probability of detection (POD). This model is used to forecast typical typhoon Dujuan (2015), and the results show that the model integrating the coverage area of extremely deep convective clouds has advantages in RI predictions, which is mainly reflected in RI forecasts when the initial intensity is strong. The model has great application potential for forecasting strong typhoons.

Based on the data of satellite altimetry and six tide gauge stations along the coast of China, linear regression function was used to estimate the absolute sea level rise rate in the coastal areas of China from 1993 to 2020, which was 4.17±1.32 mm/a, and the relative sea level rise rate was 4.47±0.90 mm/a. Taking the atmospheric data, ocean data and climate modal index from 1958 to 2020 as prediction factors, a variety of neural network models such as long short-term memory neural network model (LSTM model), recurrent neural network model (RNN model), gated recurrent unit neural network model (GRU model) and support vector machine regression model (SVR model) were established to predict the trend of relative sea level changes around the six tide gauge stations along the coast of China. The model evaluation results show that the average correlation coefficient and root mean square error of the observed value and the predicted value obtained by the LSTM model that simultaneously introduces atmospheric and ocean variables and climate modal index variables are 0.866 and 19.279 mm, respectively, which performs the best among the four models, and therefore the LSTM model can be used as a new method for predicting relative sea level changes.

The type of sea ice is one of the important attributes of polar sea ice, and the physical properties of multi-year ice are significantly different from those of first-year ice. Therefore, the identifying the types of sea ice is of great significance to the research of polar climate change and the navigation security of ships in ice-covered regions. Satellite remote sensing is an effective method to obtain multi-temporal and large-scale sea ice type information. Based on three deep learning models (ResNet, Vision Transformer, Swin Transformer) and Sentinel-1 satellite dual-polarization synthetic aperture radar (SAR) images, this paper studies the classification method for sea ice in the regions of the Northwest and Northeast Passage in the Arctic. The results showed that the sea ice classification effect of 8×8 pixel slice dataset was better than that of other size slice datasets. Offset processing false color images could effectively reduce the influence of noise on sea ice classification. Among the three deep learning models, the Swin Transformer model had the highest classification accuracy, with the overall accuracy and Kappa coefficient above 98%. Comparing the multi-year ice concentration, it was found that the results of the three models deviate less than 10% from the AMSR2 data.

The open source software OpenFOAM was used to solve the Reynolds-averaged Navier-Stokes equations, numerical simulations of wave surges (Fr=1.2~1.3) was implemented, and the results were compared with physical experimental data to verify the accuracy of the numerical simulations. The data was studied based on the CONV1D convolutional neural network model to predict the dynamic development process of tidal bore on terrain with slope. Comparison of the model prediction and numerical simulation results for the arrival time and the maximum water level of the tidal bore at x=30.0 m shows that the average relative error in the arrival time is 2.28%, and the average relative error in the maximum water level is 3.73%. The smaller relative error proves the accuracy of the CONV1D model. This result has certain research significance for the simulation of the hydrodynamic process of tidal bore, early warning of tidal bore disasters, and the study of future development trends of primary tidal bore.

Seafloor topography is of considerable value in commercial, engineering, military and scientific research. Currently, gravity field data, such as free air gravity anomalies and vertical gravity gradients, are commonly used to inverse seafloor topography. However, due to the strong multi-resolution of the existing methods to inverse seafloor topography, it is still impossible to obtain accurate high-precision seafloor topography. A new data-knowledge-driven method was proposed to reconstruct accurate seafloor topography, which combines the gravity-density method with random forests. This method was applied to the South China Sea and compared with the gravity-density, random forest, and existing SIO models. The inversion results show that the data-knowledge-driven method provides better inversion performance, followed by the random forest and gravity-density methods, and the SIO model is the worst. The mean absolute error, mean relative error and root mean square error of the data-knowledge-driven are reduced by 21%, 25% and 7%, respectively, compared to those of the gravity-density method, while they are also reduced by 20%, 20% and 20%, respectively, compared to those of the random forest. In addition, the data-knowledge-driven model has a high degree of agreement with the shipboard bathymetry data, with approximately 72% of its differences distributed within ±10 m, which is higher than that of the other three models. The results demonstrate the feasibility and effectiveness of the data-knowledge-driven approach in seafloor topography inversion, which helps to accelerate the mapping of high-precision seafloor topography.