Graph Neural Networks for Enhanced North-East Atlantic Wind Forecasting using Scatterometer Data

The rising demand for accurate, high-resolution short-term ocean forecasts requires continuous improvements of operational prediction systems. While Copernicus Marine Service Monitoring and Forecasting Centers (MFCs) are increasing their model resolution, the quality of the resulting forecast often remains constrained by inaccuracies in the operational atmospheric forcing fields. 

The Copernicus Marine Service Evolution project CERAINE addresses this issue by using data-driven correction techniques leveraging remote-sensing data. CERAINE focuses on developing Artificial Neural Networks (ANNs) to refine operational forcings (surface wind fields and surface currents) within the European North-East Atlantic (NEA) region, with a specific focus on improving surface wind fields critical for wave modeling. 

To correct systematic biases on the surface wind fields, a novel approach based on Graph Neural Networks (GNNs) is proposed. This architecture incorporates the spatial dependence of neighboring nodes, thereby inherently accounting for geographical location and context in the prediction. Furthermore, the GNN structure enforces a seamless and physically continuous correction across the entire domain, effectively eliminating blending artifacts often found in other methods. The GNN is trained using a new dataset (IFS_SC) as the target. This dataset is derived from operational wind fields from the ECMWF Integrated Forecasting System (IFS) corrected using scatterometer observations. 

Results will demonstrate the developed wind GNN performance, showcasing the benefits of the IFS_SC product over the uncorrected operational forecast. The presentation will specifically highlight how the GNN framework, leveraging the spatial coverage and accuracy of scatterometer data, significantly improves wind prediction consistency and accuracy across the entire NEA domain. Limitations and uncertainties inherent to this methodology will also be discussed.