Short-term Neural Forecasts of Sea Surface Dynamics

Short-term ocean forecasting plays a critical role in supporting marine operations, climate monitoring, and coastal management. This work introduces a modular, benchmark-driven framework for neural ocean forecasting with a specific focus on the prediction of Sea Level Height (SSH). Our benchmark suite includes an extensive set of oceanographic datasets from 2010 to 2023, integrating gridded Level 4 and gappy Level 3 satellite products, reanalysis data, and in-situ observational datasets. Neural forecasting models are trained using reproducible workflows inspired by OceanBench and assessed with standardized metrics, including normalized root mean square error (nRMSE) and derived geostrophic velocity scores. Results demonstrate the capacity of neural forecasting models to match or exceed the performance of operational forecasting systems. The entire framework, including data preprocessing, training, and evaluation pipelines, is open-sourced as part of the OceanBench-STOF project.