An AI model of the global carbon cycle

Under the name 3D Above and Below Ground Carbon Stocks (3D-ABC), Große and his team are developing a foundation model to capture and visualize the global terrestrial carbon budget, primarily from remote sensing data and large databases of biomass and soil carbon, as well as vegetation and soil carbon fluxes. From the tops of the tropical rainforests to the boreal forests of North America, from the European peat bogs to the deep permafrost soils of the Siberian tundra, the AI-based base model will initially use the vast amounts of data to capture the large terrestrial carbon reservoirs. Satellite imagery and laser data will be incorporated into the model, while multicopters and ground sensors have been providing additional relevant information for years. What's more, the resolution and power of remote sensing is increasing every year, and the observable periods are getting longer.

However, the amount of data is now too great for humans to handle alone. In addition to the AWI, the Jülich Research Centre, the German Research Centre for Geosciences, the Helmholtz Centre Dresden-Rossendorf, the Helmholtz Centre for Environmental Research and the German Aerospace Centre are also involved.

Große believes that research into permafrost soils is particularly important. After all, they are among the world's largest carbon reservoirs, currently storing around 1,600 billion tons of carbon which is almost twice as much as the atmosphere (880 billion tons). Because the Arctic is warming much faster than the global average, permafrost soils are particularly vulnerable to change. The erosion of riverbanks and coastlines, as well as the constant thawing of the landscape, exposes old, previously sequestered carbon to microbes, and ultimately to the atmosphere, exacerbating the climate crisis. Although this microbial release does not happen overnight, "it is not yet included in current climate models," says Große.

"But documenting the current state is only the beginning of the Foundation Model," says the geologist. The key question is: How will these large carbon reservoirs change in the future, in a warmer world, with even more deforestation and more intensive agriculture in certain regions? It has long been known that many land reservoirs are subject to large annual carbon fluxes. In some regions, the flux is so large that some carbon sinks become large carbon sources, at least temporarily: Instead of absorbing carbon dioxide from the atmosphere, certain ecosystems release it in large quantities. "Here, too, no climate model is currently able to simulate the sometimes very rapid developments, such as changes in land use or deforestation of rainforests, which could intensify in the coming years and decades, reliably and in detail," says Große. Artificial intelligence, which builds on the wealth of scientific observation data from satellite missions and field studies, could help.

But it is precisely this kind of information – provided in a timely, high-resolution and generally understandable form – that is essential for policymakers and business to take effective measures for climate protection and adaptation: Which sinks do we need to protect more rigorously? What sources should we avoid?

This is exactly what 3D-ABC will do. Using harmonized data from the large Landsat and Sentinel-2 satellite missions and the German radar satellite mission TanDEM-X, the basic model will first be trained to understand and reproduce the global land cover and vegetation heights in the Amazon rainforest, tundra and taiga as accurately as possible. What grows where, how high and how dense? With this information, the model will learn to determine biomass and draw detailed conclusions about the carbon budget above and below ground. Remote sensing data, for example, from satellites and drones, also record changes in the land surface. Large fires, the expansion of infrastructure, erosion or insect infestations can cause major damage to vegetation within a few days, weeks or months. This in turn affects local carbon stores. The final phase of the project will focus on carbon fluxes. "We hope to be able to train the model not only to reflect the reservoirs correctly, but also to predict carbon fluxes in high spatial and temporal detail," says Große, describing the central challenge that would be essential for actual carbon management on a global scale.

Another hurdle, according to Große, is the storage of the huge amounts of data. The information is processed at the Helmholtz Association's own supercomputer center in Jülich. However, the long-term storage capacity there is limited. A solution is currently being sought. Everyone involved in the project agrees that the potential and possible benefits of 3D-ABC are enormous. But so are the challenges that lie ahead. "To achieve the desired effect," says Große, "we have to make sure that the foundation model can work with a wide variety of imagery. With optical and radar satellite data, and with 3D information from laser measurements. If we can achieve this multimodality, we will have created an excellent tool for global monitoring of terrestrial carbon.