Imagine being able to effortlessly monitor wildlife in real time, continuously gathering information about which species live where, how they interact and how environmental changes affect them. In the midst of a biodiversity crisis — where species are disappearing before they’re even discovered — having such a capability would be groundbreaking for science, conservation and policy.
Thanks to a newly funded grant from the National Science Foundation and the Academy of Finland, Arts & Sciences Distinguished Professor of Statistical Science David Dunson and his collaborators are on a mission to build adaptable AI systems that can automatically detect which species inhabit different regions and, with the help of cutting-edge wireless technology, seamlessly transmit this data from remote locations to scientists around the world.
Traditional methods for determining species presence — manual field surveys, camera traps, environmental DNA sampling and similar approaches — are expensive, labor-intensive and time-consuming. They are also prone to errors: Some locations are far better sampled than others, and many species — particularly small, elusive or nocturnal ones — are undercounted or entirely missed. Armed only with this patchy data, scientists struggle to fully grasp how different species interact and how environmental disruptions affect these interactions.
In collaboration with a team of Finnish researchers, Dunson aims to overcome these challenges by creating cost-effective AI tools that learn and adapt on the fly. Their system integrates automated species identification using AI-enhanced analysis of audio recordings, machine-learning algorithms to filter and correct data errors, and interpretable AI models to better understand ecological dynamics.
Innovation occurs at every step of the way: From the outset, the team will build interpretable AI models to optimize data collection using Bayesian adaptive design — a decision-making method often used in clinical trials.
“Continuously recording sound is incredibly inefficient,” says Dunson. “If we’re tracking migrating birds, for example, we should focus resources on peak migration periods rather than collecting redundant data when nothing changes. Our thought is to combine statistical modeling with interpretable AI to identify the most informative locations, so we can devote more resources there.”
Once these locations have been identified, another innovation is deployed: novel 3D printed high-resolution audio-recording devices will collect sound data, which will be transmitted back to the team through cutting-edge wireless technology, eliminating the cost, time and labor typically associated with recovering data from the field.
But receiving this data is just the beginning of a new set of challenges. “The sheer scale of the data creates some major statistical and machine-learning challenges,” says Dunson. And that’s where his expertise comes in. “I’ve been leading the team that develops statistical and AI methods to solve these problems.”
One of their innovations is a new method to understand how species interact with each other and their environment.
“Let's say that we know some ecological features of a given spatial location at a certain time. Can we characterize the ecological community at that location, which species are there? And, based only on sound data, can we estimate species abundances as well?,” Dunson asks. “This indirect inference is a challenging statistical problem.”
To address it, Dunson’s team is developing new Joint Species Distribution Models — statistical models used by ecologists to explore how multiple species interact with each other and their environment — based on interpretable AI models. “Essentially all of modern AI is based on statistical modeling with deep neural networks,” he says. Most AI, like large language models, relies on more than millions of parameters that are difficult to interpret. But Dunson emphasizes that, to make sense of an ecological community, one must be able to tie a model’s parameters to real data.
“One approach we’ve pioneered is called Bayesian Pyramids,” he says. “These are multi-layer neural networks — not dissimilar from those commonly used at the foundation of AI — but we can actually constrain parameters using real data, making them more interpretable and more robust.”
Beyond developing and implementing new tools, Dunson’s team is also committed to training researchers in AI-driven ecological monitoring through workshops, open-source tools and interactive tutorials. Special emphasis will be placed on collaborating with researchers in developing regions, such as Madagascar, where biodiversity is exceptionally rich but resources for monitoring are often limited.
The methods proposed in this grant build on a successful wave of innovations developed by Dunson’s research team at Duke. He doesn’t hide his excitement: “We have really been killing it,” he says. “We have a bunch of awesome methods for this kind of joint species distribution modeling currently under revision at Nature Methods and top international statistics journals.
“This can really revolutionize ecological network modeling.”