In the present research, Indian Ocean Dipole (IOD) prediction was explored using statistical methods based on deep learning techniques. First, convolutional neural network (CNN) models were trained using sea‐surface temperature anomaly (SSTA) maps of the Indian Ocean from 1854 to 1989, and the properly trained CNN models were then validated with the period from 1991 to 2019. The results indicate that the deep learning approach is capable of forecasting the IOD at lead times up to 7 months. The forecast skills of CNN are superior to those of the dynamic models in the North American Multi‐Model Ensemble (NMME). The CNN outperforms the NMME with lower sensitivity to predictability barriers and fewer systematic errors. Moreover, the gradient heat map analysis demonstrates that the triggering precursors selected by CNN models for IOD events are novel and physically sensible. These results suggest the CNN to be a new and effective tool for both IOD prediction and comprehension.

The conditional nonlinear optimal perturbation (CNOP) approach was mainly used to investigate the effects of the uncertainties in initial condition and model parameters on model results. This study presents a new application of the CNOP approach to the investigation of the effects of boundary condition uncertainty. Specifically, we first give the general formulation of the CNOP approach for uncertainties of initial and boundary conditions and model parameters. The method is then applied to analyze the effects of nutrient perturbations at the bottom boundary of the water column on the modeled deep chlorophyll maximum (DCM) using an ocean ecosystem model. The results show that nutrient perturbations at the bottom boundary have significant impacts on the modeled DCM. Interestingly, the approach also reveals that nonlinear processes play important roles in the evolution of phytoplankton perturbations caused by nutrient perturbations at the bottom boundary. This implies that the CNOP approach is useful for investigating the effects of boundary condition uncertainty.

Using the Regional Ocean Modeling System and the conditional nonlinear optimal perturbation (CNOP) approach, this study explored a CNOP‐based adaptive observation system design with regard to improving the prediction of one seasonal upstream Kuroshio transport (UKT) reduction event. For comparison, investigations were also conducted using a set of random observation systems built within a sensitive area inferred from previous studies and a conventional observation system built at 18°N. For these observation systems, observation networks constructed using different numbers of observation sites and observation distances (i.e., the shortest distances between any two observation sites) were evaluated by performing observing system simulation experiments. Results showed that most assimilation experiments can improve the UKT prediction, indicating the importance of initial conditions. For all observation systems, large numbers of observation sites generally led to higher prediction benefits, whereas the effects of observation distances differed significantly. Overall, the CNOP‐based observation system exhibited the best performance. Optimal CNOP‐based observation networks were established using six or eight observation sites and observation distances of 140 or 165 km, which produced a mean prediction improvement of 43.1%. This preliminary work is expected to provide guidance on future observations in the upstream Kuroshio region and to help realistic UKT predictions.

The Kuroshio Intrusion (KI) into the South China Sea (SCS) is a key process that greatly modulates the variability of dominant phenomena in the northern SCS. Using the Regional Ocean Modeling System, the optimal precursor (OPR) triggering the KI is obtained by the Conditional Nonlinear Optimal Perturbation (CNOP) approach. The OPRs exhibit the anticyclonic eddies (AEs) spatial pattern in the southern region (19°N–20.3°N, 119°E–122°E) of the Luzon Strait. The evolution processes of the OPRs indicate that the northwestward movement and rapid growth of AEs force the Kuroshio to bend westward continuously, which leads to the transition of the Kuroshio path from leaping state to intruding state. Eddy‐energetic analyses suggested that the barotropic instabilities are responsible for the evolution of the OPRs and thus triggering the KI, while the local wind stress plays a minor role in the triggering process. Additional examinations based on satellite altimeter data and ocean reanalysis data further confirm that the AE‐like perturbation similar to the obtained OPR indeed plays a key role in the KI triggering process.