If a hiker gets lost in the rugged Scottish Highlands, rescue teams sometimes send up a drone to search for clues of the individual’s route—trampled vegetation, dropped clothing, food wrappers. But with vast terrain to cover and limited battery life, picking the right area to search is critical.

Traditionally, expert drone pilots use a combination of intuition and statistical “search theory”—a strategy with roots in World War II-era hunting of German submarines—to prioritize certain search locations over others. Jan-Hendrik Ewers and a team from the University of Glasgow recently set out to see if a machine learning system could do better.

Ewers grew up skiing and hiking in the Highlands, giving him a clear idea of the complicated challenges involved in rescue operations there. “There wasn’t much to do growing up, other than spending time outdoors or sitting in front of my computer,” he says. “I ended up doing a lot of both.”

To start, Ewers took datasets of search and rescue cases from around the world, which include details such as an individual’s age, whether they were hunting, horseback riding or hiking, and if they suffered from dementia, along with information about the location the person was eventually found—by water, buildings, open ground, trees, or roads. He trained an AI model with this data, in addition to geographical data from Scotland. The model runs millions of simulations to reveal the routes a missing person would be most likely to take under their unique circumstances. The result is a probability distribution—a heat map of sorts—indicating the priority search areas. 

With this kind of probability map, the team showed that deep learning techniques could be used to design more efficient search paths for drones. In research published last week on arXiv, which has not yet been peer reviewed, the team tested its algorithm against two common search patterns: the “lawnmower,” in which a drone would fly over a target area in a series of simple stripes, as well as an algorithm similar to Ewers’ but less adept at working with probability distribution maps.

In virtual testing, Ewers’ algorithm beat both of those approaches in two key measures; the distance a drone would have to fly to locate the missing person, and the percentage of time the person was found. While the lawnmower and existing algorithmic approach found the person 8% of the time and 12% of the time, respectively, Ewers’ approach found them 19% of the time. If it proves successful in real rescue situations, the new system could speed up response times, and save more lives, in scenarios where every minute counts. 

“The search and rescue domain in Scotland is extremely varied, and also quite dangerous,” Ewers says. Emergencies can arise in thick forests on the Isle of Arran, the steep mountains and slopes around the Cairngorm Plateau, or the faces of Ben Nevis, one of the most revered but dangerous rock climbing destinations in Scotland. “Being able to send up a drone and efficiently search with it could potentially save lives.”

Search and rescue experts say that using deep learning to design more efficient drone routes could help locate missing persons faster in a variety of wilderness areas, depending on how well suited the environment is for drone exploration (it’s harder for drones to explore dense canopy than open brush, for example).

“That approach in the Scottish Highlands certainly sounds like a viable one, particularly in the early stages of search when you’re waiting for other people to show up,” says David Kovar, a director at the US National Association for Search and Rescue in Williamsburg, Virginia, who has used drones for everything from disaster response in California to wilderness search missions in New Hampshire’s White Mountains. 

But there are caveats. The success of such a planning algorithm will hinge on how accurate the probability maps are. Overreliance on these maps could mean that drone operators spend too much time searching the wrong areas. 

Ewers said a key next step to making the probability maps as accurate as possible will be obtaining more training data. To do that, he hopes to use GPS data from more recent rescue operations to run simulations, essentially helping his model to understand the connections between the location where someone was last seen and where they were ultimately found. 

Not all rescue operations contain rich enough data for him to work with, however. “We have this problem in search and rescue where the training data is extremely sparse, and we know from machine learning that we want a lot of high quality data,” Ewers says. “If an algorithm doesn’t perform better than a human, you are potentially risking someone’s life.”

Drones are becoming more common in the world of search and rescue. But they are still a relatively new technology, and regulations surrounding their use are still in flux.

In the US, for example, drone pilots are required to have a constant line of sight between them and their drone. In Scotland, meanwhile, operators aren’t permitted to be more than 500 meters away from their drone. These rules are meant to prevent accidents, such as a drone falling and endangering people, but in rescue settings such rules severely curtail ground rescuers’ ability to survey for clues. 

“Oftentimes we’re facing a regulatory problem rather than a technical problem,” Kovar says. “Drones are capable of doing far more than we’re allowed to use them for.”

Ewers hopes that models like his might one day expand the capabilities of drones even more. For now, he is in conversation with the Police Scotland Air Support Unit to see what it would take to test and deploy his system in real-world settings.