Could geospatial foundation models help improve conservation effectiveness?

This blog post was written by Professor Julia Jones, Bangor University and co-chair of the Chief Scientists’ Group of the Joint Nature Conservation Committee. This is the first in a series of blog posts from our “Improving Conservation Evidence” meeting, held in January 2026.

On the first day of the Improving Conservation Evidence meeting we brought together the science leads of all five UK Statutory Nature Conservation Bodies (SNCBs) to discuss how advances in artificial intelligence (AI) can contribute to their work. We were joined by Professor Anjali Goswami (DEFRA’s chief scientist), experts from the University of Cambridge Department of Computer Science and Technology, and colleagues from the Conservation Evidence Group.

One of the aims of this workshop was to give the science leads of the SNCBs the opportunity to learn about recent developments in geospatial foundation models, such as Cambridge’s TESSERA model. These models are likely to transform how we integrate ground-based ecological data with earth observation data in future. This could have dramatic implications for how we map habitat extent and condition. evaluate past interventions, and shape future conservation policy and practice. 

We opened the workshop with a brief update from each of the SNCBs (the Joint Nature Conservation Committee, Natural England, Natural Resources Wales, NatureScot, and the Northern Ireland Environment Agency) sharing how they are currently incorporating AI into their practice. It was a valuable opportunity to share experiences. The discussion was practically focused and reflected nuanced consideration of ethical issues and trust around the use of AI.

Dr Sadiq Jaffer then took the floor to present the University of Cambridge’s open-source geospatial foundation model: TESSERA.

We have vast, and ever-growing, quantities of earth observation data available. However, making useful maps requires appropriate ground-truthed training data, technical skills, and a lot of computing power. The availability of all three can be real bottlenecks to making the most of earth observation.

TESSERA overcomes these bottlenecks. The team have ingested a huge amount of satellite data from multiple passes a month over most of the earth at 10 m resolution (thanks to the EU’s Copernicus programme). Their model works out how each pixel differs from other pixels, and simplifies this down to an annual time series of ‘embeddings’. These abstract away cloud cover while retaining signals coming from seasonal changes.

The TESSERA ‘embeddings’ are rich in ecological meaning (TESSERA embeddings are outperforming its competitor proprietary models when tested against a range of tasks from predicting tree height to below ground fungal diversity). The richness of these embeddings means that even a relatively small amount of training data can be used to build a map. One of the team, Professor Anil Madhavapeddy, did a demo for us by building a (very basic) habitat map from just a few data points of training data. The model was run locally on his laptop emphasising another advantage of working with these embeddings rather than raw satellite data.

There is justifiable concern about the energy costs of large AI models and the resulting impact of the growth in datacentres on net zero targets. In the case of TESSERA, while ingesting the data to create the embeddings is costly, this energy use is shared by all the end users. The end users run simple models rather than processing heavy satellite data themselves which saves energy. The usefulness, ease of use, and the fact they are completely free to use means the number of users will grow rapidly, bringing down the ‘per use’ cost.

The presentation resulted in a fascinating discussion about how such innovations may influence the work of the SNCBs in future. The next step is to train and validate downstream models combining existing ground-based data with the embeddings. The hope is that this would allow us to interpolate datasets between sampled locations and then ask larger-scale questions about ecosystem change and the impact of interventions for a wider range of outcomes than is currently possible. Such understanding is crucial for so many applications.

Dr Sara McGuckin, from the Northern Ireland Environment Agency (NIEA), said:

“In Northern Ireland, we have rich data on a range of environmental variables and are keen to explore how emerging technology can help further unlock the power of such data to continue to improve the effectiveness of our conservation management and environmental decision-making.  I welcome the opportunity for continued engagement with such a vibrant and innovative conservation evidence community”.

The second day of the meeting was a conference which brought together different communities from across conservation to discuss the challenge of improving conservation effectiveness. Despite the freezing cold (the heating wasn’t working in our venue – a beautiful converted church), the atmosphere was warm and we all learned a huge amount.

Professor Sallie Bailey, the Chief Scientist of Natural England, said:

“The conference offered a rare opportunity for funders, academics, policy-makers, policy advisors and the science leaders from arms-length bodies to consider the critical role of evidence in driving transformative change for nature. This is so valuable to ensure evidence developed is directly aligned to policy and practice need – so we can all act effectively with urgent ambition”.

Thanks to the Conservation Evidence Group, the Cambridge Centre for Science and Policy, and Pembroke College, Cambridge for hosting this incredibly valuable meeting, and to Professor Bill Sutherland, from the University of Cambridge Department of Zoology, for making it happen.