Oceanographers have developed a dual-architecture deep learning system called Eddy LSIN that substantially improves the accuracy of mesoscale eddy trajectory forecasts in the northeastern South China Sea. Mesoscale eddies are rotating oceanic vortices that function as critical weather-scale dynamic processes in marine environments, typically spanning spatial scales of roughly one hundred kilometers and persisting across intervals ranging from days to tens of days. Because these circulating water masses govern horizontal and vertical water column movement, energy transfer, and dynamic ocean stratification, tracking and forecasting their paths remains essential for maritime navigation, climate change research, marine environmental monitoring, and sustainable resource management. However, predicting eddy tracks has historically presented severe computational hurdles due to the coupled, highly nonlinear physical mechanisms driving their life cycles, regional and seasonal variability, and the rapid accumulation of forecasting errors in traditional numerical modeling approaches.
To overcome the limitations of classic hydrodynamic models and earlier machine learning implementations, the research team constructed the Eddy LSIN framework by uniting a long short-term memory recurrent neural network for short-range forecasting with an enhanced temporal Informer model, designated T-Informer, for extended horizons. Standard numerical simulations based on classical ocean dynamic equations frequently require complex parameterizations that struggle in fine-resolution settings or in topographically intricate marine regions. Prior deep learning efforts utilizing recurrent or convolutional architectures similarly encountered obstacles, largely because they relied on iterative multi-step forecasting where single-step outputs feed back into the model as subsequent inputs. In long-sequence simulations, this recursive dependency causes temporal errors to compound rapidly, degrading forecasting precision past several days while failing to integrate crucial physical features and seasonal variations.
A foundational contribution of the new framework is the construction of a high-dimensional multimodal dataset named Spatiotemporal Eddy Representation, or STER, designed to capture the intrinsic physical state of circulating water masses. The researchers assembled historical eddy tracking records from the Global Mesoscale Eddy Trajectory Dataset covering the multi-decade period from January 1, 1993, to December 31, 2020, in the northeastern South China Sea within coordinates fifteen to twenty-two degrees North and one hundred eleven to one hundred twenty-four degrees East. Across this domain, they extracted core trajectory attributes including center longitude and latitude, tracking speed, effective radius, amplitude, observation time, and eddy polarity. To provide comprehensive dynamic context, the authors integrated daily gridded sea level anomaly and geostrophic current data at one-quarter degree resolution alongside reanalysis sea surface temperature fields downscaled from one-twelfth degree resolution.
Rather than relying solely on raw sea surface heights and velocity fields, the investigators explicitly derived relative vorticity and eddy kinetic energy across surrounding twelve-by-twelve grid patches, which cover an area of approximately 330 by 330 kilometers centered on each observed eddy. These spatial dimensions were selected because empirical analysis in the northeastern South China Sea confirmed that over ninety percent of cyclonic and anticyclonic eddies exhibit diameters under 300 kilometers. By calculating relative vorticity from spatial velocity gradients and deriving kinetic energy from geostrophic velocity anomalies, the team captured the fundamental rotation and kinetic intensity of each structure. These features were arranged into a four-layer three-dimensional matrix of size four by twelve by twelve, which was subsequently compressed using a convolutional autoencoder and principal component analysis into a compact ten-dimensional representation, reducing data redundancy while preserving coherent physical dynamics.
In comparative evaluations testing the impact of this feature processing pipeline, the STER dataset demonstrated marked advantages over a standard unprocessed baseline dataset comprising sea level anomaly, temperature, and horizontal velocity components. When evaluated using identical network configurations and hyperparameters, the model trained on the autoencoder-derived STER representations achieved a rapid, smooth convergence without the severe validation error fluctuations observed with the baseline dataset. On the reserved test split, the mean squared error dropped to 0.0020 and the mean absolute error reached 0.0301 when utilizing STER inputs. Compared to the baseline configuration’s test errors of 0.0670 mean squared error and 0.1746 mean absolute error, the STER representation yielded an approximate seventy percent reduction in mean squared error and an eighty-two percent reduction in mean absolute error.
For the short-term prediction phase spanning lead times from one to seven days, the Eddy LSIN framework relies on a long short-term memory network configured with 128 hidden units and trained with a batch size of 32. Across one-day forecast intervals, this network achieved an average center displacement error of approximately 8.12 kilometers, and maintained an average error of 18.524 kilometers at the seven-day forecast mark. These results demonstrated strong spatial overlap between predicted coordinates and observed satellite tracks across most of the northeastern South China Sea. However, the researchers noted that when the LSTM architecture was pushed beyond one week using standard step-by-step iterative forecasting, its tracking performance deteriorated markedly, with average spatial displacement errors compounding to 66.114 kilometers by day twenty.
To sustain trajectory accuracy across extended prediction intervals extending from seven to twenty days, the researchers introduced the T-Informer model, adapting the transformer-based Informer architecture to mitigate cumulative iterative drift. Traditional transformer attention models often struggle with extended temporal dependencies in environmental sequences, but the T-Informer leverages sparse self-attention and distillation operations to generate multi-step prediction sequences directly rather than iteratively. Crucially, the authors incorporated a seasonal time-encoding mechanism tailored to the cyclical variations of the South China Sea. Because regional eddy kinetic energy and relative vorticity exhibit pronounced annual cycles—peaking strongly during the winter months from December to February and diminishing between April and October—the T-Informer embeds temporal records into continuous dense vectors representing four distinct seasons, fusing them with positional encodings to capture periodic atmospheric and oceanic forcing.
To properly optimize geographical accuracy alongside numerical regression fidelity, the study formulated a combined mean squared error objective function denoted as CMSE. Traditional neural network loss functions rely entirely on Cartesian coordinate differences, which can fail to represent true spatial distances across spherical geographic coordinates. The CMSE loss explicitly combines standard mean squared error with a spherical Haversine distance metric through an adjustable weighting parameter lambda. Sensitivity analysis revealed that purely numerical optimization where lambda equals one caused average spatial prediction errors to rise to 8.7681 kilometers on one-day forecasts, whereas setting lambda to zero overly prioritized geographic distance at the cost of numerical stability. The authors determined that setting lambda to 0.7 yielded optimal performance, minimizing one-day distance error to 8.1207 kilometers while maintaining stable training dynamics.
When evaluated across multi-day benchmark tests, the T-Informer model consistently outperformed standard and specialized sequence-learning architectures over long lead times. At day twenty, the T-Informer attained an average center positioning error of 39.24 kilometers, representing an approximate forty percent error reduction compared to conventional sequence models and substantially outperforming the 66.114-kilometer error observed with the recursive LSTM. Furthermore, test evaluations on individual tracks showed average coordinate errors remaining as low as 0.1583 degrees longitude and 0.1442 degrees latitude at the twenty-day forecast horizon.
The researchers also investigated the spatial transferability of the Eddy LSIN framework by applying the trained system directly to the broader South China Sea domain spanning ten to twenty-four degrees North and one hundred eight to one hundred twenty-seven degrees East over the 2018 to 2020 period without regional retraining. In this cross-domain evaluation, the framework yielded average geographical errors of 10.31 kilometers at one day, 23.78 kilometers at seven days, 30.18 kilometers at ten days, 41.80 kilometers at fifteen days, and 53.88 kilometers at twenty days. While these regional errors were moderately higher than those recorded in the native northeastern training sector due to differing circulation patterns and bathymetry, the steady, degradation-free performance confirmed that the framework captured generalized eddy dynamics rather than overfitting to localized bathymetric artifacts.
Despite its demonstrated forecasting precision, the authors documented important regional limitations, particularly in the vicinity of the Luzon Strait where trajectory errors were noticeably elevated. The Luzon Strait serves as a primary conduit for Kuroshio current intrusion, intense water-mass exchanges, and energetic submesoscale turbulence, creating complex nonlinear flow fields that challenge satellite observation resolution and model representation. Moreover, the current Eddy LSIN implementation operates purely as an end-to-end data-driven network lacking explicit physical constraints such as geostrophic balance, conservation laws, or energy budgets. The authors concluded that future developments must focus on integrating multi-source in situ data from buoys and research vessels, testing cross-basin transferability in systems such as the Gulf Stream or equatorial Pacific, and incorporating physics-informed neural network architectures to ensure dynamical consistency in topographically complex straits.
Subject of Research: Deep learning models for mesoscale ocean eddy trajectory prediction in the northeastern South China Sea
Article Title: Mesoscale eddy trajectory prediction in the Northeastern South China Sea based on an deep learning framework
Article References: Han, T., Song, D., Li, C., Zhao, W., & Wang, X. (2025). Mesoscale eddy trajectory prediction in the Northeastern South China Sea based on an deep learning framework. Intelligent Marine Technology and Systems, 3(1), Article 26. https://doi.org/10.1007/s44295-025-00077-0
Image Credits: AI Generated
DOI: 10.1007/s44295-025-00077-0
Keywords: Mesoscale Eddies, Deep Learning, South China Sea, Ocean Trajectory Prediction, Eddy LSIN, LSTM, T-Informer, Physical Oceanography, Satellite Altimetry, Sea Level Anomaly, Eddy Kinetic Energy, Relative Vorticity, Autoencoder, CMSE Loss, Predictive Oceanography, Ocean Dynamics, Time Series Modeling
Cite Scienmag News
Blake Davidson. (November 21, 2026). Deep Learning Framework Improves Mesoscale Ocean Eddy Trajectory Forecasts. Scienmag. https://scienmag.com/deep-learning-framework-improves-mesoscale-ocean-eddy-trajectory-forecasts/
Blake Davidson. "Deep Learning Framework Improves Mesoscale Ocean Eddy Trajectory Forecasts." Scienmag, 21 November 2026, https://scienmag.com/deep-learning-framework-improves-mesoscale-ocean-eddy-trajectory-forecasts/. Accessed 21 November 2026.
Blake Davidson. "Deep Learning Framework Improves Mesoscale Ocean Eddy Trajectory Forecasts." Scienmag. November 21, 2026. https://scienmag.com/deep-learning-framework-improves-mesoscale-ocean-eddy-trajectory-forecasts/








