Plymouth Marine Laboratory works with IBM to launch new AI model to monitor ocean health

Plankton bloom in the Barents Sea | European Space Agency

Among the first-of-its-kind, the new ‘Granite-Geospatial-Ocean’ model represents a significant advance in the capacity to monitor marine ecosystem health and carbon cycling from space.

It joins a growing family of AI models developed by IBM Research to help better understand Planet Earth and its climate.

The Ocean remains challenging to monitor comprehensively due to its dynamic and multi-dimensional nature. Conditions at sea for data collection can also be harsh or inaccessible, and direct measurement campaigns are expensive and spatially limited, similar to measuring a single pixel on a cinema screen.

Granite-Geospatial-Ocean has been developed to help address these issues by optimising and learning from both remote sensing data alongside carefully calibrated, directly sampled measurements.

The state-of-the-art model analyses satellite imagery and in situ marine measurements to provide rich, high-quality insights into ocean colour, phytoplankton distributions and net primary production, the rate at which sunlight is converted into energy by photosynthesising organisms such as phytoplankton.

The model builds on IBM’s vision transformer architecture (derived from their Prithvi model) and adapted for marine environments. It was pre-trained on approximately 500,000 satellite images from ESA’s Copernicus Sentinel-3 missions, and then fine-tuned on a selection of high-quality in situ observations.

At one tenth the size of some land/Earth system models, it has been engineered to be efficient and flexible, making it accessible to a wide range of users.

To demonstrate what the model can do, the project team developed two immediate ‘downstream’ applications using IBM’s open-source TerraTorch toolkit:

• Phytoplankton distribution estimation: mapping how microscopic marine plants are spread in the sunlit ocean surface.
• Net primary production estimation: calculating how much organic carbon phytoplankton generate via photosynthesis.

In tests over the Atlantic and coastal waters off Iberia, Granite-Geospatial-Ocean outperformed classical machine learning baselines in reproducing ground-truth chlorophyll concentrations and spatial productivity patterns, even though only a comparatively small number of field measurements were used for tuning.

Phytoplankton are vital to life not just in the ocean but also life on this planet so they are vital to understand. However, they are difficult to monitor at the level of detail needed for accurate marine models.

Dr David Moffat, project lead and AI & Data Science Lead for PML, said:

“Phytoplankton are taken wherever the currents may go. It’s as if we’re chasing around forests to understand how they’re producing food for marine life.”

“This new model is a step change for what we could potentially accomplish with remote sensing in marine science”.

These successes suggest the model could be extended further, for instance to detect harmful algal blooms, monitor sediment and nutrient run-off or track other biogeochemical processes.

PML scientists contributed field measurements, domain expertise and validation guidance to ensure the model performs robustly in real-world marine settings.

Both Granite-Geospatial-Ocean and its downstream applications are being openly released via Hugging Face, following the project’s commitment to transparency, accessibility and scientific collaboration.

By providing a scalable, adaptable tool to map ocean colour and productivity globally, this initiative has the potential to refine estimates of how much carbon the oceans absorb and store; an important uncertainty to tackle in climate modelling.

In an IBM Research blog post Anne Jones, a senior research scientist at IBM working on climate-related technologies, said:

“The foundation model approach could help to refine our primary production estimates and give us a better sense of how much warmer the planet could get in the decades ahead”.