For decades, ecology has been very good at explaining the past. We can look at a collapsed fishery, a vanished pollinator population, or a coral reef bleaching event and piece together, after the fact, what went wrong. What ecology has struggled to do, and what it needs to get better at, is predicting these changes before they happen.
Conservation policymakers don’t just need to understand why an ecosystem collapsed last year. They need to know what will happen if we build a road through a forest, dam a river, or let a new pesticide into widespread use, before those decisions are made, not after the damage is done.
A new method paper by C. J. M. Musters and Geert R. de Snoo of Leiden University, published in Individual-based Ecology, explores whether combining deep learning with this individual-based framework could offer a way forward. They propose: closing the prediction gap requires not just better data, but an entirely different kind of data – and possibly, machines that sense the world the way animals do.
Why Predicting Ecosystems Is So Hard
Ecological communities are dazzlingly complex. Every organism in a habitat is constantly responding to its environment and to every other organism around it, like competing for food, avoiding predators, seeking mates, adjusting behavior with the seasons. Traditional ecology often simplifies this complexity by grouping organisms into species-level categories.
Musters and de Snoo point to the work of ecologist John Lawton, who argued that this kind of unpredictability is strongest at “intermediate” scales – not in small populations of a single species, not across whole biomes, but somewhere in between. The authors define this middle ground as spaces of roughly 1 to 1,000,000 km² over 10 to 100 years. Unhelpfully, that’s also roughly the scale at which conservation decisions actually get made.
Organism-based Ecology or Individual-based Ecology takes a different starting point. Instead of assuming species behave as uniform units, it treats individual organisms as the fundamental actors whose actions and interactions generate the patterns we see in ecosystems. That’s a more realistic picture – but also a far more complicated one. Communities become vast networks of individual organisms, each constantly gathering and responding to information from their surroundings and from each other.
Where Deep Learning Comes In
Deep learning (DL), the technology behind everything from image recognition to language models, is extraordinarily good at finding patterns in complex data. In principle, DL combined with organism-based ecology, the study of how individual organisms behave and interact, could be used to build models that predict how a community will shift as conditions change.
However, deep learning is only as good as the data you feed it, and the data ecology currently collects wasn’t designed with this goal in mind.
Traditional ecological datasets tend to record outcomes: population counts, species distributions, migration patterns. What they largely fail to capture is something more fundamental – the raw sensory information that organisms actually use to make decisions. A bird doesn’t respond to “average temperature over the season.” It responds to what it can see, smell, hear, and feel, moment to moment, in its environment.
If deep learning models are going to predict how organisms actually behave and interact, the authors argue, they need training data that reflects what the organisms themselves are sensing rather than just retrospective statistics ecologists usually collect.
Building “Artificial Sense Organs” for Ecology
To generate that new kind of data, Musters and de Snoo propose building sensory instruments that mimic the sense organs of the organisms themselves – broad-spectrum devices that continuously and non-invasively record the same signals animals, plants, fungi, and microbes use to make decisions. Paired with existing tools like camera traps, biologging, satellite imagery, and environmental DNA sampling, this could generate the rich, organism-centered datasets that deep learning models actually need.
A Call for Caution
Deep Learning models are black boxes, even when their predictions are accurate, it’s often unclear why. Models trained on incomplete or biased data can generate ecologically impossible results, like population sizes that go negative.
Also, models trained largely on data from wealthy, well-studied regions could embed cultural assumptions that don’t transfer elsewhere. The authors offer a vivid example, Black-tailed Godwits are cherished in the Netherlands but considered agricultural pests in parts of West Africa, where they feed on rice rather than soil invertebrates. Applying a European-trained model globally, without care, risks a kind of “unintentional colonialism” in conservation decision-making.
In their method paper, the authors are clear, that the models they describe don’t exist yet, and that building them will require new instruments, new interdisciplinary collaboration between ecologists and ethologists, and critically explicit, published testing of where these models work and where they don’t. What they’re offering instead is a roadmap, by starting small, building sensory prototypes, training experimental models on well-studied plots, and rigorously checking every prediction against reality before trusting it with real conservation decisions.
Whether “artificial sense organs” turn out to be the right way to get there remains to be seen. But the underlying challenge – teaching machines to understand ecosystems well enough to predict their future – is one of the more important questions facing conservation science today.
Original source:
Musters CJM, de Snoo GR (2026) Challenges and opportunities for Artificial Intelligence for predicting changes in ecological communities. Individual-based Ecology 2: e199762. https://doi.org/10.3897/ibe.2.199762
