satellite imagery. However, conventional modeling methods rely on users to manually engineer predictive features from these datasets—for example generating slope angles and curvatures from terrain elevation data in the hope that these can help explain the spatial distribution of the variable being mapped.

However, scientists believe there are likely to be many more nuanced relationships at play within the natural environment that models based on traditional manual feature-engineering approaches may simply miss.

The pioneering new AI approach, developed in the study, poses environmental information extraction as an optimization problem. Doing so allows it to automatically recognize and make use of relationships that may otherwise go unnoticed and unutilized by humans using more traditional modeling methods.

In addition to improving map quality, this approach also unlocks the potential for the discovery of new relationships in the natural environment by AI, while simultaneously eliminating huge amounts of trial-and-error experimentation in the modeling process.

Charlie Kirkwood, a postgraduate student at the University of Exeter said: “To be useful for decision making, we need our models to provide answers that are as specific as possible while also being trustworthy—and that means creating accurate measures of the uncertainty associated with our estimates, which in this case are predictions at unmeasured locations.”

“Our AI approach is set within a Bayesian statistical framework, which allows us to quantify these uncertainties and provide a range of uncertainty measures, including credible intervals, exceedance probabilities and other more bespoke products that will feed directly into decision making processes. Crucially, all this is provided whilst harnessing any available information more effectively than traditional approaches allow—which you can see coming through in the detail of the map.”