Enabling Ocean AI: Foundation Models in WEkEO Workspace

WEkEO’s JupyterHub GPU-enabled environment supports the effective use of foundation models for ocean science by making fine-tuning accessible, scalable, and practical. By allowing users to adapt pre-trained models to their own data and use cases, it helps bridge the gap between experimentation and real-world research and application. 

Satellite remote sensing remains the only viable way to observe the global ocean surface continuously at high spatial and temporal resolution. At the same time, artificial intelligence (AI) foundation models (FMs) have transformed how Earth observation data can be exploited. Yet in ocean science, the bottleneck has not been the models themselves, but the ability for users to adapt and operationalise them efficiently. 

Foundation models are powerful because they are pre-trained on vast volumes of unlabelled satellite data, learning general trends or dynamics for some variables such as ocean colour. However, their real value emerges only when they are fine-tuned for specific scientific or operational use cases, a step that requires both computational power and accessible environments. This is precisely where WEkEO plays a decisive role. 

From Access to Adaptation 

WEkEO’s graphical processing unit (GPU)-enabled JupyterHub environment is designed to move users beyond experimentation into practical, scalable fine-tuning of foundation models. 

Rather than requiring users to build and maintain their own high-performance infrastructure, WEkEO provides a cloud-based workspace where large satellite datasets and GPU resources are directly accessible.  

Users can accelerate large-scale numerical computations, run inference on foundation models, and test machine-learning prototypes before scaling them up. The environment supports Python, R, RStudio, and Julia, reflecting the diverse practices of the ocean-science community and enables multiple users to work in parallel with dedicated GPU memory.  

The IBM Ocean Foundation Model example 

To illustrate these capabilities, IBM Research UK has developed an open-source ocean foundation model trained on Copernicus Sentinel-3 data, including ocean colour and sea-surface temperature observations from the OLCI and SLSTR instruments. 

This model benefits from consistent, global-scale records that capture both physical and biogeochemical ocean processes. As a pre-trained model, it provides a strong baseline representation of ocean dynamics. 

However, its real utility lies in what users can do next: fine-tune it within the WEkEO workspace to address their own scientific needs. 

From General Models to Targeted Applications 

Inside the WEkEO environment, users can take a general-purpose foundation model and tailor it to specific applications. Typical examples include estimating chlorophyll-a concentrations, a key indicator of phytoplankton biomass and deriving primary production, central to understanding the marine carbon cycle. Through fine-tuning, the model can incorporate local datasets, improve performance in specific regions, and better capture temporal variability. 

Crucially, this workflow is not limited to a single model or use case. The same approach can be applied to a wide range of challenges, from coastal to ecosystem change monitoring. 

Scaling Impact Through Accessible AI 

The combination of foundation models and WEkEO’s GPU allows: 

• Faster transition from research to applications  
• Greater reproducibility through shared environments and datasets  
• Lower barriers for institutions without dedicated HPC resources  
• Stronger community collaboration around open models and workflows  

WEkEO’s JupyterHub GPU-enabled environment provides a space where users can take existing foundation models and make them their own. In a context where the ocean is undergoing rapid and unprecedented change, this capability is not just technical; it is strategic. It enables the ocean community to move from data access to data-driven decision-making, powered by adaptable, scalable AI.

Image: Primary production estimation with the Ocean Foundation Model for Sentinel-3. Credit: European Union, WEkEO, Information (2025), ©Mercator Ocean.