SEA-ViT: How a Transformer Reads Ocean Currents ✛
Context
Forecasting surface currents is critical for navigation, fisheries management, and climate modeling. Traditional deterministic models struggle to capture the nonlinear spatiotemporal dependencies of ocean dynamics. This work aligns with the application of deep learning to geospatial data, alongside research on explainable AI, self-supervised learning, and radar image processing. However, SEA-ViT stands out by directly predicting vector components of currents with climate index input, rather than classifying or segmenting images.
Methods
The authors compiled a 30-year high-frequency radar dataset, including timestamps, coordinates, velocity components U and V, and a categorical ENSO index (0 for neutral, 1 for El Niño, 2 for La Niña). Data were split into training, validation, and test sets in a 70/15/15 ratio. U and V were normalized using z-scores. To improve robustness, augmentation was applied: small random perturbations to coordinates and Gaussian noise to U and V. The architecture consists of a bidirectional GRU processing the temporal sequence and a transformer self-attention layer modeling spatial correlations. The loss function is mean squared error with an additional penalty for errors during El Niño and La Niña periods. Training and deployment are automated via an MLOps pipeline, with a Swagger UI for inference.
Results
SEA-ViT demonstrated accurate forecasting of surface current components U and V in the Gulf of Thailand and the Andaman Sea. Including the ENSO index improved predictions during climate anomalies. Thanks to transformer self-attention, the model captures long-range spatial dependencies that convolutional networks with limited receptive fields cannot. The bidirectional GRU effectively accounts for both short-term tidal fluctuations and seasonal trends. Quantitative accuracy metrics are not provided in the preprint, but the authors emphasize the model's practical applicability for operational forecasting.
Implications
SEA-ViT shows that hybrid transformer-recurrent architectures can model complex spatiotemporal covariance of ocean currents. Unlike neighboring works that often tackle classification or image segmentation, this work addresses regression of a vector field with external climate forcing, expanding the toolkit of geoinformatics and oceanography and offering an alternative to computationally intensive numerical models. Practical value is underscored by plans to integrate into the GISTDA system.
Future development
In the coming months, the authors plan to fully integrate SEA-ViT into GISTDA's AI infrastructure. Three stages are planned: data preparation and augmentation with additional variables, iterative model optimization with hyperparameter tuning, and deployment and evaluation in real conditions. Further development may include wind forcing, bathymetry, and satellite data. The model could also be applied to other regions and adapted for sea surface temperature forecasting.
Impact
The model could improve marine navigation, fisheries management, environmental monitoring, and climate analysis in Thai coastal zones. Integration into GISTDA systems would provide operational current forecasts for government and commercial users.
Next steps
Immediate next steps include training the model on the full 30-year dataset, fine-tuning hyperparameters, and deploying in GISTDA's production environment. Subsequent validation on independent observations and comparison with traditional numerical models are planned.
Key open problems
This work addresses a fundamental oceanography problem: predicting nonlinear current dynamics under climatic variability. Open questions remain about model transferability to regions with different oceanography and its ability to capture rare extreme events. Interpretability is also unresolved: which spatiotemporal patterns does the transformer extract?
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The paper proposes the SEA-ViT model for predicting flow velocity components (U, V) in Thai coastal waters. The architecture combines a Vision Transformer for spatial dependencies and bidirectional GRUs for temporal dynamics, with an additional feature: the ENSO index. The model was trained on 30 years of high-frequency radar data and demonstrated high accuracy; the code and weights are publicly available.
Where its strength is
The strength of the work lies in the hybrid architecture, which captures long-range spatial correlations through self-attention while also accounting for both short-term and seasonal temporal patterns via bidirectional GRUs. Incorporating the ENSO index as an external driver improves forecasting during periods of climate anomalies.
Where it could go next
- It might be interesting to look at explainability methods from the weather AI paper to highlight which spatiotemporal patterns the model deems significant for current forecasting. This would help interpret the contribution of the ENSO index and self-attention mechanisms.based on: Peeking Inside the Mind of a Weather AI arXiv:2604.20467
- There is a related approach for predicting El Niño up to one and a half years ahead; one could consider using longer-term climate predictors or pretrained ENSO embeddings to improve forecast lead time for currents.based on: El Niño Can Be Predicted a Year and a Half in Advance arXiv:2601.12251
- The archive contains a paper on a neural network that, by reading weather backwards, warns of droughts months in advance; its scheme of processing sequences in reverse could be useful for strengthening the temporal component of SEA-ViT, especially for capturing slow oceanic processes.based on: How a neural network, reading weather backwards, warns of drought months in advance arXiv:2607.00331
Worth trying
- Test the explainability techniques from [2604.20467] to visualize transformer attention maps across different ENSO phases.
- Try adding long-term climate features to the model, as in [2601.12251], and check whether accuracy improves on horizons beyond one week.
- Use reverse time-series processing, as in [2607.00331], as an additional GRU pass and compare metrics.
Why this may matter
If SEA-ViT can be scaled to other coastal areas with minimal fine-tuning, it could provide a rapid tool for operational current forecasting, which is critical for navigation and responding to climate change.
Nearby works in our archive
- Radar Sculptor: How Radio Waves Sculpt a Drone's Portraitsimilarity 0.52
- Peeking Inside the Mind of a Weather AIsimilarity 0.51
- Automatic Galaxy Classification from Radio Data: The STRADAViT Methodsimilarity 0.50
- Gravitational Tango: How a Neural Network with Physical Intuition Heard the Rhythm of Black Holessimilarity 0.50
- An Underwater Hunter for Earth's Heatsimilarity 0.50
- Cellular Echoes: How Your Phone Navigates Without GPSsimilarity 0.50
- El Niño Can Be Predicted a Year and a Half in Advancesimilarity 0.49
- How a neural network, reading weather backwards, warns of drought months in advancesimilarity 0.49
The archive is sparse around this work: the closest review is only 0.52, against a usual 0.64. It looks like few have drilled nearby.
The model is named SEA-ViT, simultaneously referencing the sea and the Vision Transformer architecture. The author thanks himself in the acknowledgments section, a rarity in scientific papers.
Key numbers
- data duration: 30 years
- data split: 70% training, 15% validation, 15% test
- current vector dimension: 2 components (U, V)
- ENSO index values: 0 neutral, 1 El Niño, 2 La Niña