Ge, Linyao, Wilson, Richard orcid.org/0000-0001-7265-3033, Zhang, Xinmin et al. (2 more authors) (Accepted: 2026) TLKNet: Mesoscale Eddy Trajectory Prediction with Topology Aware Large Kernel Convolutional Neural Network. IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING. ISSN: 0196-2892 (In Press)
Abstract
Mesoscale eddies play a crucial role in oceanic energy generation and dissipation, with their trajectories reflecting energy transport and cascade processes. Accurate trajectory prediction is essential for both ecological and economic applications. However, most existing prediction algorithms focus solely on eddy attributes and local environmental features, overlooking the oceanic topological structure, which largely shapes the dynamical and morphological parameters of eddies. Therefore, in this paper, we construct a topology-aware large kernel convolutional neural network, named TLKNet, which improves the accuracy of topological structure prediction and thereby enhances the accuracy of mesoscale eddy trajectory prediction. Its convolution kernel is designed based on mesoscale signal processing that empowers the model’s mesoscale feature representation. Compared with popular deep learning methods, the proposed model achieves a significant improvement in performance. Specifically, TLKNet achieves a multi-day averaged geodetic distance error of 7.8 km on the test set and improves the accuracy of mediumrange prediction by approximately 20%. Meanwhile, the eddy trajectory prediction problem is transferred from the instancelevel to the gridded-field method.
Metadata
| Item Type: | Article |
|---|---|
| Authors/Creators: |
|
| Dates: |
|
| Institution: | The University of York |
| Academic Units: | The University of York > Faculty of Sciences (York) > Computer Science (York) |
| Date Deposited: | 12 Nov 2026 11:58 |
| Last Modified: | 12 Nov 2026 11:58 |
| Status: | In Press |
| Refereed: | Yes |
| Open Archives Initiative ID (OAI ID): | oai:eprints.whiterose.ac.uk:247930 |

CORE (COnnecting REpositories)
CORE (COnnecting REpositories)