Nature's patterns, unlocked with AI.

Smart software could predict the next invasive plant species or disease outbreak.

ECOLOGISTS have it hard. The ecosystems they study often contain dozens if not hundreds of interacting species, each one affected by myriad variables such as weather, soils, predators and local history. In such a jungle, it can be hard to find a path to understanding. But now ecologists have a new guide: artificial intelligence.

Already in use in disciplines such as stock market analysis, software that can recognise hidden patterns and trends in messy data offers a powerful tool for ecology.

“The real strength of machine learning is you can take these really complex databases that contain tonnes of species interacting in ways we can't even imagine, and let the data speak for themselves,” says Barbara Han, an ecologist at the University of Georgia, Athens, who co-organised a symposium on machine learning in ecology at last week's meeting of the Ecological Society of America in Austin, Texas.

The strength of AI is that you can take complex ecosystems and let the data speak for themselves

For example, ecologists have struggled for decades to predict which introduced plant species are likely to spread and become serious pests, but so far they haven't managed to find a method that works. So John Paul Schmidt, also at the University of Georgia, turned to AI.

First, Schmidt gathered data on the size, habitat and other biological attributes of nearly 7900 horticultural plants introduced into Hawaii, as well as their pest status. Then he ran the data through an AI process known as boosted regression tree analysis, which makes repeated guesses about which factors might predict pest status, gradually refining its choices to end up with the strongest predictors.

The results surprised him. “When I started out, I thought things like growth form – shrub versus herbaceous versus vine – were going to be good explanatory variables,” he says. But instead, he found that one of the best predictors of a potential pest is a large number of chromosomes, especially compared with related species (PLoS One, DOI: 10.1371/journal.pone.0017391). Plant species with more chromosomes have probably experienced a doubling of chromosome numbers more recently in their evolutionary history, which may give them more genetic raw material for adapting to their new environments, Schmidt says.

Besides their ability to find subtle patterns, AI techniques have two big advantages over conventional statistical analyses, says Schmidt. First, they do not assume, as most conventional analyses do, that measurements such as seed size or chromosome number follow a bell-shaped distribution – an assumption that is often violated in the real world. Second, the AI techniques cope easily with missing data, such as species for which seed size or chromosome number are unknown, whereas conventional analyses often omit such species from the analysis.

AI techniques do have their problems, though, not least that they require lots of data. If too little data is available, they can “overfit” – that is, they make far more precise predictions than are warranted, much as someone looking at just two weeks' weather data might conclude that Fridays are the warmest days simply because those two Fridays were.

This was shown by Reuben Keller, an ecologist at the University of Chicago in Illinois, who found that AI techniques performed no better than conventional statistical methods at predicting invasive species of birds, fish, molluscs or pine trees when data sets contained just 18 to 87 species (Diversity and Distributions, DOI: 10.1111/j.1472-4642.2011.00748.x).

There are now many examples of AI's use in ecology. Han is just beginning a project that will use AI to predict where emerging diseases will pop up, and which animal species will transmit them (see “Spotting the next SARS”) . John Drake of the University of Georgia is using AI to identify the most important interactions among species within an ecosystem. And Bill Langford of the Royal Melbourne Institute of Technology in Australia is using it to help set conservation priorities more effectively.

Viruses that jump to humans from other animals often make front-page news: just look at influenza, Ebola virus, SARS coronavirus and HIV. Public-health workers spend a lot of time and money watching for potential epidemics to emerge, and trying to control them when they do.

“Wouldn't it be great if we knew where to look in advance,” says Barbara Han of the University of Georgia in Atlanta. Han hopes that her new project will help to achieve that. The idea is to use AI to sort through three sorts of data simultaneously.

Han plans to analyse the characteristics of wild animals – their size, metabolic rate, habitat, diet and so forth – to try to understand why some animals are good reservoirs for diseases. She will also examine the characteristics of viruses to work out which ones are likely to spread to humans. Finally, she will study the characteristics of habitats with the intent of predicting where viruses are likely to make the leap.

AI can also help us make sense of the behaviour of individual animals. It is already being used to identify individual penguins in video images. Now Robin Freeman, a computational ecologist at Microsoft Research in Cambridge, UK, and his colleagues are using it to identify particular behaviours in migratory seabirds.

Freeman's team has tagged individual Manx shearwaters (Puffinus puffinus) with geolocators that allow them to track the bird's location each day. They also collect a continuous record of whether the leg bearing the tag is immersed in water, which indicates that the bird is either sitting on the surface or diving to feed.

The team used a machine-learning method to tease from the immersion record three discrete behavioural patterns – one with little immersion that corresponded to migratory flight and two corresponding to feeding on winter and summer ranges. When they matched these sorted records to the location data, they found that the birds made mid-ocean feeding stops during their migration from the UK to South America – something no one had been able to prove before (Proceedings of the Royal Society B, DOI: 10.1098/rspb.2008.1577).

Such simple behavioural analysis is only the beginning, says Freeman. As more and more researchers deploy instruments like these, they should be able to discriminate animal behaviours more finely using, for example, accelerometers to identify feeding dives. He speculates that it may even be possible to design instruments that recognise when an animal is behaving unusually, and then adapt by collecting extra data about that new behaviour.

“It feels like we're at the threshold of another field, the informatics of behaviour in wild animals,” Freeman says. “It's exciting to think what we'll be able to do 10 years from now.”