Modern anti-personnel mines are small and often made with plastic casings that standard metal detectors cannot detect. Geophysical techniques such as ground-penetrating radar, magnetometry and electromagnetic induction are significantly less effective at detecting plastic mines than metal ones.
Particular concern surrounds so-called scatterable mines, which are designed to be deployed across large areas. One of the most widely used is the Soviet-era PFM-1 mine, commonly known as the "Petal" or "Butterfly" because of its distinctive shape.
"Caring for a wounded soldier is more difficult than dealing with a dead one. These mines are designed to wound, not kill. They are specifically engineered for that purpose, and their entire design is intended to make them difficult to detect," said Alex Nikulin, an associate professor in the Department of Earth Sciences at Binghamton University.
The new approach uses machine learning algorithms, a form of artificial intelligence, to detect plastic mines across large areas.
In active war zones such as Ukraine, scatterable mines often remain close to the surface. In post-conflict areas, however, they may gradually become buried or concealed by the landscape, said Sharifa Karvandyar, a Binghamton University geology graduate. Because plastic mines are roughly the size of a mobile phone, drones used for detection must fly at low altitudes — about 10 to 20 meters (33 to 66 feet) above the ground — to maximize sensor resolution.
"This is a first-pass analysis to determine whether an area is potentially hazardous," Karvandyar said. "It aligns with the standard landmine detection process."
For her master's thesis, Karvandyar used a drone-mounted camera and image-stitching software to create low-resolution aerial images. Those images were then processed using the You Only Look Once (YOLO) machine learning algorithm to identify potential mines.
The researchers trained the YOLO object-detection model using inert PFM-1 mines and 3D-printed replicas. They placed the mines throughout the Binghamton University Nature Preserve to create a dataset showing how PFM-1 mines appear in different environments, from various angles and under different lighting conditions.
"We trained two different YOLO models to see how we could make this approach field-ready," Karvandyar said. "One model was trained only on PFM-1 mines, while the other was trained to identify both PFM-1 mines and additional random objects using a standard dataset."
The second model delivered lower performance, which Karvandyar said likely better reflects real-world conditions because cameras also capture environmental features such as leaves.
Most of the data processing takes place during the algorithm's training phase, which lasts from several hours to a full day depending on the number of images. Once trained, the system requires only a standard laptop, a drone and a camera for field deployment.
That is a significant advantage because both active conflict zones and post-war oblasts often suffer from poor internet connectivity, either because infrastructure has been damaged or because GPS and communications signals are heavily jammed, as is the case in Ukraine. To be practical in the field, the system must operate without an internet connection, which this method allows.
Earlier, NV Tech reported that in 2024, Carnegie Mellon University doctoral student Mateo Dulce Rubio developed an AI-powered mine detection system. He led the development of RELand, a tool designed to help humanitarian organizations more accurately identify mine-contaminated areas.
Developed by a team of students and researchers, RELand uses artificial intelligence to analyze mine contamination data. It combines machine learning with geographic and sociodemographic data to produce more accurate predictions of landmine contamination risk.