OceanBench - Short-Term Global Ocean Forecasting

The increasing adoption of AI-based approaches in Earth system sciences has led to breakthroughs in modeling and forecasting, exemplified by state-of-the-art performance of neural weather forecasting systems [Bi et al., 2023, Lam et al., 2022]. In oceanography, Deep Learning techniques show significant promise for advancing ocean state modeling by combining both modeled and observational ocean datasets [Febvre et al., 2023, Martin et al., 2023, Wang et al., 2024]. However, in the context of ocean forecasting, the deployment of neural forecasting approaches faces challenges such as sparse observational data and uncertainties in existing datasets. Despite advances in ocean observation systems, the ocean remains under-sampled, complicating the training of robust forecasting models [Wang et al., 2024].

This study presents the application of the 4DVarNet framework [Fablet et al., 2021, Fablet et al., 2023] in forecast mode, specifically for 7-day sea surface height (SSH) prediction. 4DVarNet employs an end-to-end Deep Learning strategy to forecast future SSH state from sparse satellite observational data. Using a variational data assimilation formulation, the framework combines a UNet with a convolutional LSTM to iteratively reconstruct future ocean state. The model was trained on synthetic altimetry observations sampled from the GLORYS12 operational reanalysis (2010–2019) and evaluated on independent Nadir altimetry tracks from 2023.

The results demonstrate that 4DVarNet outperforms traditional state-of-the-art operational forecasting system GLO12 [Lellouche et al., 2013], achieving a normalized RMSE (nRMSE) score of 0.92 for lead time 0 compared to 0.86 for the baseline GLO12. The model shows superior accuracy across all forecast lead times, highlighting its potential advantage in operational oceanography. The framework improves the accuracy of the SSH forecast by effectively using gappy satellite altimetry data. This demonstrates a particular interest of applying the proposed method to other data sources, such as the SWOT altimetry mission, but also of implementing alternative learning strategies, including training on synthetic datasets and fine-tuning with real-world altimetry observations.

In addition to the improved predictive performance compared to the state-of-the-art operational forecast system, this study establishes a standardized workflow for data processing, training, and evaluation, inspired by OceanBench framework [Johnson et al., 2023].

This research highlights the potential of the 4DVarNet framework for short-term neural ocean forecasting. The proposed method efficiently handles sparse altimetry data and achieves significant performance in predicting 7-days gap-free SSH state at global scale, improving the accuracy by almost 65% in average compared to the baseline GLO12. Future studies should focus on improving the model by incorporating additional data sources, evaluating the impact of input and output resolutions and associated learning strategies (eg. patching), and exploring its applicability to additional variables describing ocean state other than the SSH.