AI helps map how ecosystems absorb CO₂ across Europe
New research by Deltares and Universities of Delft, Twente, Leipzig and Vienna shows how artificial intelligence can help monitor the health of ecosystems and their role in the carbon cycle across different landscapes -without the need for site-specific calibration.
Understanding how much carbon ecosystems absorb is essential in the fight against climate change. Forests, grasslands, and agricultural systems play a major role in capturing CO₂, yet measuring this uptake consistently and at a large scale remains challenging. The study explores how data-driven models can help bridge this gap. By combining field measurements with satellite data, researchers demonstrate how artificial intelligence (AI) can be used to estimate ecosystem productivity across different locations in Europe.
The article has been published in Elsevier Ecological Informatics
From local measurements to large-scale insights
Ecosystem productivity is traditionally measured using specialised flux towers. While these provide highly accurate data, they are limited in number and only represent specific locations.
To overcome this limitation, researchers combined flux tower measurements with satellite observations from the European Sentinel-2 mission. They then applied three modelling approaches—statistical models, machine learning and deep learning—to estimate how much carbon ecosystems absorb over time.
The results show that modern machine learning methods can successfully upscale local measurements to regional level, offering new opportunities for large-scale environmental monitoring.
One model, multiple ecosystems
A key innovation of the research is the development of a cross-site modelling approach. Rather than building a separate model for each location, the researchers investigated whether a single model can be applied across different ecosystems—from forests to grasslands and cropland.
The study shows that this is indeed feasible. The best-performing model mantained good accuracy even when applied to sites it had not seen before.
“This is an important step towards monitoring ecosystem productivity at larger scales, without the need for detailed local calibration,” says Anna Spinosa, data scientist ecosystem health monitoring at Deltares and first author of the article.
Different strengths for different models
The study also compared different modelling techniques:
- Machine learning (XGBoost) showed the most consistent performance across sites
- Deep learning (LSTM) was better at capturing extreme events, such as peak productivity
- Statistical models remained useful as a baseline but struggled with complex patterns
This highlights that no single method captures all aspects of ecosystem behaviour. Combining approaches may offer the best results in future applications.
Implications for climate and ecosystem management
By enabling scalable monitoring, these methods could support decision-making at national and European level.
The findings are relevant for a wide range of applications:
- Climate policy: better estimates of carbon uptake support more reliable carbon accounting
- Nature-based solutions: improved monitoring of forests and wetlands helps assess their impact
- Water and land management: insights into ecosystem productivity support adaptation to drought and climate change
Challenges and next steps
While the results are promising, the study also highlights important limitations. Model performance varies between ecosystems, and models can still struggle with highly dynamic environments such as agricultural systems.
Future work will focus on improving model robustness by incorporating additional data, such as soil properties and new satellite products, and by expanding the approach to more regions.
Towards operational ecosystem monitoring
Overall, the research demonstrates the potential of combining satellite data and AI for monitoring ecosystem functioning at scale.
As Earth observation datasets grow and AI methods improve, this approach could form the basis for operational systems that track ecosystem health and carbon uptake in near real time.
