The researchers on the project, from Binghamton University in New York, had previously developed a method that allowed for highly accurate detection of 'butterfly' landmines using low-cost commercial drones equipped with infrared cameras.

This new research, however, focuses on automated detection of landmines using convolutional neural networks, the standard machine learning method for object detection and classification in the field of remote sensing. According to Alek Nikulin, an assistant professor of energy geophysics at the university, this method is a “game-changer” in the field.

“All our previous efforts relied on human-eye scanning of the dataset,” Nikulin explained. “Rapid drone-assisted mapping and automated detection of scatterable minefields would assist in addressing the deadly legacy of widespread use of small scatterable landmines in recent armed conflicts and allow to develop a functional framework to effectively address their possible future use.”

It is estimated that there are at least 100 million military munitions and explosives of concern devices in the world, of various size, shape and composition. Millions of these are surface plastic landmines with low-pressure triggers, such as the mass-produced Soviet PFM-1 'butterfly' landmine.

Nicknamed for their small size and butterfly-like shape, experts have said these mines are extremely difficult to locate and clear due to their small size, low trigger mass and, most significantly, a design that mostly excluded metal components, making these devices virtually invisible to metal detectors.

More than a million of these mines have littered Afghanistan since military helicopters dropped them across the country during the Soviet-Afghan War in the 1980s. Their green shell and winged shape often attracts children to play with them, earning them grim notoriety as “the toy mine.”

In their paper on the subject, the researchers wrote that they used convolutional neural network (CNN)‐based approaches to automate the detection and mapping of landmines. They also highlighted the method’s importance.

They wrote: “One, it is much faster than manually counting landmines from an orthoimage (i.e. an aerial image that has been geometrically corrected). Two, it is quantitative and reproducible, unlike subjective human‐error‐prone ocular detection. And three, CNN‐based methods are easily generalisable to detect and map any objects with distinct sizes and shapes from any remotely sensed raster images.”

The researchers tested the system on thermal datasets of butterfly mines collected by drones, using the neural networks to search the images for light emitted by the mines. They claimed the system yielded a 99.3 per cent testing accuracy for a partially withheld testing set and a 71.5 per cent testing accuracy for a completely withheld testing set.

This approach, according to the team at Binghamton, is much faster than manually counting landmines. It is also reproducible, they said, unlike the subjective and error-prone human methods of detection.

The researchers believe that these detection and mapping techniques are generalisable and transferable to other munitions and explosives of concern. For example, they could be adapted to detect and map disturbed soil for improvised explosive devices (IEDs).