IBM and partners open-source a new AI model for monitoring Earth’s oceans

Oceans cover two-thirds of our planet, and even today, they remain underexplored. Harsh weather, powerful currents, and murky waters that increase in pressure with depth, have kept all but the most adventurous away. But now, a new porthole for discovery has opened — without having to spend weeks at sea.

Working in collaboration with the Plymouth Marine Lab (PML), UK Science and Technology Facilities Council (STFC) Hartree Centre, and University of Exeter, IBM Research has released a first-of-its kind foundation model for the ocean, called Granite-Geospatial-Ocean.

The model takes the architecture that IBM Research created for Prithvi, the pioneering AI model for analyzing geospatial data, and applies it to the oceans. Trained on raw satellite images and fine-tuned on physical measurements gathered at sea, this new color-coded map of the world’s oceans could provide new insights into the health of marine ecosystems and their role in cycling carbon through the biosphere.

Granite-Geospatial-Ocean joins a growing family of AI models developed by IBM Research to better understand our planet, its climate, and the star at the center of our solar system. To validate the model and motivate others to try it, the researchers built two applications using TerraTorch, an open-source framework designed by IBM Research to simplify the process of fine-tuning geospatial models.

The first application can be used to estimate the distribution of phytoplankton, tiny, plant-like microorganisms living in the oceans’ sunlit upper layer. The other can calculate net primary production, the amount of organic matter that phytoplankton collectively incorporate into their bodies through photosynthesis.

The biomass that phytoplankton produce by harvesting the Sun’s energy feeds virtually all marine life, from zooplankton that eat the phytoplankton, to the fish, birds, sharks, and whales higher up the food chain.

Phytoplankton also indirectly help to regulate climate by taking up inorganic carbon from the air through photosynthesis and transferring a portion of it to the deep ocean, where it can stay buried for thousands of years. How much carbon gets stored long-term, however, is unclear, and this uncertainty carries over to climate prediction models.

The new ocean model could help to bring both processes into clearer focus. "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,” said Anne Jones, a senior research scientist at IBM working on climate-related technologies.

Both Granite-Geospatial-Ocean and its downstream applications are available on Hugging Face, joining the IBM-NASA Prithvi models for monitoring weather and climate and changes over land, and the IBM-European Space Agency (ESA)’s multimodal TerraMind models, which also focus on land-based change.

Granite-Geospatial-Ocean borrows Prithvi’s vision transformer architecture and was trained in a similar way by having it reconstruct partially blacked-out satellite images to teach it how to fill in spatial gaps in the data. It’s the first foundation model to be trained on color images captured by the ESA’s Copernicus Sentinel-3 satellite. At 50 million parameters, a tenth the size of Prithvi-EO-2.0, it’s also relatively quick and easy to run and customize on a range of devices.

A moveable feast

Phytoplankton are as vital to life in the oceans as plants and trees are to land — just tricker to pin down. Each drop of seawater in the top hundred meters of the ocean contains thousands of these unicellular organisms.

“They're taken wherever the currents may go,” said David Moffatt, an AI researcher at PML who helped build the ocean model. “It’s as if we're chasing around forests to understand how they’re producing all this food for other marine life.”

A pigment called chlorophyll-a allows phytoplankton to convert sunlight and carbon dioxide into chemical energy that fuels their growth. Oxygen is released in the process, helping to make Earth hospitable to life. Scientists estimate that phytoplankton are responsible for producing nearly half of the oxygen in Earth’s atmosphere.

Phytoplankton incorporate about 50 billion tons of inorganic carbon into their cells each year through photosynthesis— about as much as plants and trees do on land. About 20% of this carbon, by one estimate, eventually falls to the deep ocean where it’s effectively locked away until deep currents resurface it.

The oceans have absorbed at least a third of the carbon dioxide humans have pumped into the air since industrialization, and this impressive feat is due largely to phytoplankton. Whether phytoplankton can keep up this pace in a warmer climate, however, is uncertain. A recent study analyzing 25 years of ocean color data found that net primary production had declined in tropical and sub-tropical waters, likely because of reduced nutrient upwelling triggered by warmer surface waters.

The view from space

Much of our detailed knowledge about phytoplankton’s role in cycling carbon and other elements through the biosphere comes from measurements collected on scientific research ‘cruises’ and from buoys. Despite the high quality of the data, its sparsity has made it difficult to make global estimates of primary production and carbon uptake.

In recent decades, satellites have brought a broader perspective. They create color-coded maps of the ocean’s upper layer by bouncing light off the oceans’ surface and measuring the wavelength bands scattered back. Bright green hues point to more chlorophyll-a, and by extension, more phytoplankton and energy to feed the food web; darker blues indicate less productivity.

Granite-Geospatial-Ocean was pre-trained on about 500,000 color-coded images and fine-tuned on a minimal amount of high-quality field data corresponding to exact dates in the satellite footage.

In a new pre-print study, the research team showed that the model could reproduce ground-truth chlorophyll-a concentrations in the Atlantic Ocean more accurately than a classical machine-learning model trained on identical data. They found similar results reconstructing primary production rates off the coasts of Portugal and Spain, an area known for its rich biodiversity.

In both cases, the foundation model was able to produce more accurate spatial patterns across large areas of the ocean. The results even surprised the researchers, given how little field data they used to tune the model (about 100 and 200 measurements, respectively). With more observational data, estimates of phytoplankton abundance and primary production could be extended to the rest of the oceans, and be used to refine climate change estimates, they said.

High-quality observations of other ocean processes could be used to adapt Granite-Geospatial-Ocean to many more environmental monitoring tasks, from detecting harmful algae blooms (HABs), which can threaten fisheries and public health, to tracking runoff sediment and nutrients, which can degrade water quality.

“I see this as a step change for what we could potentially accomplish with remote sensing in marine science,” said Moffat, at PML.