Finding and removing land mines is an excruciatingly slow process. Human deminers scour contaminated ground inch by inch with handheld metal detectors, waiting for the telltale beep of a magnetic anomaly. Although trained dogs are sometimes used, metal detectors have remained the go-to clearance method since the end of World War II. “There’s a very long period where there hasn’t been much innovation in the field,” says Jasper Baur, a Ph.D. student in volcanology and remote sensing at Columbia University. Baur and his collaborators at Safe Pro Group, a manufacturer of personal protective gear, have been developing a drone-based machine-learning technology to make demining safer and faster than with traditional methods.
The idea is deceptively simple: A drone flies over an area thought to be mined, collecting a large volume of images. Baur’s algorithm, trained on the visual characteristics of 70 types of land mines, cluster munitions, and other unexploded ordnance, processes the images into a map, with resolution down to a fraction of an inch. The model can then recognize and map explosives more quickly and accurately than a human reviewing the same images. “In a matter of minutes you’ll have a map plotted out with where all the land-mine detections are,” Baur says.
For now the AI can detect only surface-level explosives, not deeply buried ones or those covered by vegetation. Baur’s nonprofit organization, the Demining Research Community, is testing ways to look deeper by using thermal imaging and ground-penetrating radar. It is also developing a model that can rate the AI’s level of confidence in its mine-detection results based on the amount of vegetation present.