What I learnt at the National Academy of Sciences US-UK Forum on Biodiversity / Jun 2025

I spent a couple of days at the National Academy of Sciences in the USA at the invitation of the Royal Society, who held a forum on "Measuring Biodiversity for Addressing the Global Crisis". It was a packed program for those working in evidence-driven conservation:

Assessing biodiversity is fundamental to understanding the distribution of biodiversity, the changes that are occurring and, crucially, the effectiveness of actions to address the ongoing biodiversity crisis. Such assessments face multiple challenges, not least the great complexity of natural systems, but also a lack of standardized approaches to measurement, a plethora of measurement technologies with their own strengths and weaknesses, and different data needs depending on the purpose for which the information is being gathered.

Other sectors have faced similar challenges, and the forum will look to learn from these precedents with a view to building momentum toward standardized methods for using environmental monitoring technologies, including new technologies, for particular purposes. -- NAS/Royal Society US-UK Scientific Forum on Measuring Biodiversity

I was honoured to talk about our work on using AI to "connect the dots" between disparate data like the academic literature and remote observations at scale. But before that, here's some of the bigger picture stuff I learnt...

Shifting conservation to a winning stance

The need for urgent, additional action came across loud and clear from all the top actors in biodiversity. On the bright side, we have made stellar progress in measuring more dimensions of biodiversity accurately than ever before in human history. But, the field of biodiversity does not have a single "simple question" that needs answering, unlike many other science challenges in physics or chemistry. The ecosystem of nature measurements need to span scales ranging from the micro (from fungi and soil health) to the macro (species richness and diversity), with geographical coverage across the planet but also hyperlocal accuracy for ecosystem services.

One key question asked at the forum was how we can get to interoperable, pragmatic tools that enable all the actors involved in conservation actions (from the governments that set policy, to the private sector that controls the supply chains, to the people who have to live in and depend on natural services) to work together more effectively on gathering all the data needed.

This interoperability has to emerge during a rapid shift towards digital methods, which are vulnerable to being deleted and edited at scale with decades of painstaking observations at risk at the moment. And in the middle of all this, machine learning is swooping in to perform data interpolation at scale, but also risks dividing and polluting observations with inaccurate projections.

What is an optimistic future for conservation?

This is all quite the challenge even for a gung-ho computer scientist like me, and I was struggling with the enormity of it all! But things really clicked into place after the inspirational Julia P.G. Jones pointed me at a fantastic big-picture paper:

Drawing reasonable inferences from current patterns, we can predict that 100 years from now, the Earth could be inhabited by between 6-8 billion people, with very few remaining in extreme poverty, most living in towns and cities, and nearly all participating in a technologically driven, interconnected market economy.

[...] we articulate a theory of social–environmental change that describes the simultaneous and interacting effects of urban lifestyles on fertility, poverty alleviation, and ideation.

From Bottleneck to Breakthrough: Urbanization and the Future of Biodiversity Conservation

They observe that the field of conservation has often "succumbed to jeremiad, bickering, and despair". Much of this angst springs from the (failed) bets made by Paul Ehlrich, who thinks humans are going to be wiped out because of unbounded expansion. In response, conservation has become "the art of slowing declines" rather than achieving long term wins. But instead of being moribund, the paper paints an optimistic, practical endgame for conservation:

We suggest that lasting conservation success can best be realized when:

• the human population stabilizes and begins to decrease
• extreme poverty is alleviated
• the majority of the world's people and institutions act on a shared belief that it is in their best interest to care for rather than destroy the natural bases of life on Earth.

It turns out that most of these conditions can be reasonably projected to happen in the next fifty years or so. Population is projected to peak by the turn of the century, extreme poverty might reasonably be eradicated by 2050, and urban landuse will stabilise at 6% of terrestrial land by 2030-ish.

Given this projection, the paper then points out that conservation doesn't need to save nature "forever". Instead, we have to save enough nature now to "breakthrough" from the great acceleration of WWII until we stabilise landuse.

The profound danger is that by the time the foundations of recovery are in place, little of wildlife and wild places will be left. If society focuses only on economic development and technological innovation as a mechanism to pass through the bottleneck as fast as possible, then what remains of nature could well be sacrificed. If society were to focus only on limiting economic growth to protect nature, then terrible poverty and population growth could overwhelm what remains.

Either extreme risks narrowing the bottleneck to such an extent that our world passes through without its tigers, elephants, rainforests, coral reefs, or a life-sustaining climate. Therefore, the only sensible path for conservation is to continue its efforts to protect biodiversity while engaging in cities to build the foundations for a lasting recovery of nature. -- From Bottleneck to Breakthrough

This puts what we need to achieve today in a far, far more pragmatic light:

[...] it means that conservation faces another 30–50 years of extreme difficulty, when more losses can be expected. However, if we can sustain enough nature through the bottleneck—despite climate change, growth in the population and economy, and urban expansion—then we can see the future of nature in a dramatically more positive light.

Conservation is all about solving difficult opportunity-cost decisions in society. Science can help calculate credible counterfactuals that allow policymakers to balance limited resources to minimise nature harm while maximising benefit to humans. We can also figure out new economic methods to figure out the value of future actions. When combined, this can help conservation break through the bottleneck of the next fifty years of nature loss... and computer science can make a serious accelerative impact here (yay!).

Topics relevant to our planetary computing research

Having got my existential big-picture crisis under control, here are some more concrete thoughts about some of the joint ideas that emerged from the NAS meeting.

Resilience in biodiversity data

We've been doing a lot of work on mechanisms to process and ingest remote sensing data. All of our techniques also apply to biodiversity, except that the pipelines are even more complex due to the multi-modal nature of the data being stored. This can be clearly seen in this review on the decline of insect biodiversity that speaker Nick Isaac and my colleague Lynn Dicks published last month.

The data itself isn't just from one source; instead, we need a pipeline of spatial (at different resolution) measurements, of different types (visual, acoustic, occurrence), of different provenance (experts, crowdsourced, museum), and from different hypotheses tests (evidence bases).

Once the ingestion pipeline is in place, there's a full range of validation and combination and extrapolation involved, often involving AI methods these days. The output from all of this is then tested to determine which conservation actions to take.

Andrew Gonzalez also talked about the ambitious global biodiversity observing system that he's been assembling a coalition for in recent years. They are using Docker as part of this via their Bon in a Box product but hitting scaling issues (a common problem due to the size of geospatial tiles).

There's a good tie in for collaboration with us here via the next-generation time-travelling shell that Patrick Ferris is developing that can handle this via ZFS snapshots. Michael Dales has been applying this to scaling the LIFE and FOOD pipelines recently with Alison Eyres and Thomas Ball. And meanwhile Josh Millar and Hamed Haddadi have been researching embedded biodiversity sensors. The overall theme is that we need to make the hardware and software stack involved far easier to use for non-expert programmers.

Observing the earth through geospatial foundation models

Another problem that several speakers discussed was how complex biodiversity observations are to manage since they span multiple scales. In my talk, I described the new TESSERA geospatial foundation model that Frank Feng, Srinivasan Keshav and Sadiq Jaffer have been leading in Cambridge. As this is a pre-trained foundation model, it needs to be finetuned to specific downstream tasks. A number of people came up after my talk with suggestions for collaborations here!

Firstly, Kat Bruce (fresh from spraying pondwater with Prince William) explained how NatureMetrics are gathering eDNA from many diverse sources. The data is of varying licenses depending on which customer paid for the acquisition, but overall there is a lot of information about species presence that's very orthogonal to the kind of data gathered from satellite observations.

Secondly, Barnabas Daru from Stanford described his efforts to map plant traits to species distribution models. This complements some work David A Coomes has been leading recently in our group with Ian Ondo and Neil Burgess on mapping rare plants globally. The basic problem here is that plant occurrence data is extremely data deficient and spatially biased for 100k+ species, and so we'll need cunning interpolation techniques to fill in the data gaps.

When back in Cambridge, I'm going to arrange for all of us to chat to see if we can somehow combine eDNA, fungal biodiversity, plant traits and satellite foundation models into a comprehensive global plant species map!

Evidence synthesis from the literature

There was also huge enthusiasm for another of our projects on analysing the academic literature at scale. While we've been using it initially to accelerate the efficiacy and accuracy of systematic reviews for Conservation Evidence, there are a huge number of followup benefits for having a comprehensive data corpus.

Firstly, Chris Elphick pointed out a metasynthesis where they manually integrate recent hypotheses about insect stressors and responses into a network (3385 edges / 108 nodes). It found that the network is highly interconnected, with agricultural intensification often identified as a root cause for insect decline. Much like the CE manually labeled dataset, it should be possible to do hypothesis searches in our LLM pipeline to expand this search and make it more dynamic.

Secondly, Oisin Mac Aodha, fresh from a recent talk in Cambridge, discussed his recent work on few-shot species range estimation and also WildSAT text/image encoding. His example showed how you could not only spot a species from images, but also use text prompts to refine the search. An obvious extension for us to have a go at here is to combine our large corpus of academic papers with these models to see how good the search/range estimation could get with a much larger corpus of data.

And thirdly, I finally met my coauthor David Williams in the flesh for the first time! We've worked together recently on the biodiversity impact of food, and we had a long discussion over dinner about whether we could glean more behavioural data about how people react from the wider literature. This would require us expanding our literature corpus into grey literature and policy documents, but this is something that Sadiq Jaffer and I want to do soon anyway.

The connective tissue across these seemingly disparate projects is that there is a strong connection between what you can observe from space (the canopies of trees) to the traits expressed via knowledge of plant physiology and their DNA. If we could figure out how to connect the dots between the observed species to the physiological traits to the bioclimatic range variables, we could figure out where the (many) data-deficient plant species in the world are! I'll be hosting a meeting in Cambridge soon on this since we're already working on it.

Visualisations in biodiversity

The most unexpectedly cool talk was Ron Milo showing us visualisations of the mass distribution of all life on earth. His work really puts our overall challenge into context, as it shows just how utterly dominated wildlife is by domesticated animals.

It struck me just how important these sort of high-level visualisations are in putting detailed numbers into context. For example, he also broke down global biomass that showed that plants are by far the "heaviest" living thing on earth, and that the ocean organisms do still dominate animal biomass.

My favourite new animation library on the block is AnimeJS, and so once I plan to try to do some nice animations for LIFE and FOOD along these lines after the academic term finishes.

And that's a wrap on my notes for now! I'm still hanging out in the US for a bunch more meetings (including one at National Geographic HQ), so I'll update this note when the official RS/NAS videos and writeup comes out.

(Update 5th June: the full talk videos series is now online at the National Academy of Sciences channel. Enjoy!)