AI-powered drones are now scanning minefields in Ukraine at a pace that manual survey teams cannot match, with one company reporting that its detection dataset has surpassed 36,000 landmine identifications drawn from more than 2,000,000 analyzed drone images. The technology pairs aerial sensors with deep learning algorithms to spot explosive remnants that are nearly invisible to the human eye, particularly small plastic mines designed to evade traditional metal detectors. As researchers and defense firms race to scale these systems, the results so far suggest a dramatic shift in how contaminated land can be cleared and returned to civilian use.
36,000 Detections Across Ukrainian Minefields
Safe Pro Group, Inc. announced that its AI dataset has now logged over 36,000 landmine detections across approximately 22,500 acres of contaminated terrain in Ukraine. The company processed more than 2,000,000 drone images to reach that figure, representing one of the largest publicly reported operational datasets for AI-assisted mine detection. Those numbers matter because they move the conversation beyond laboratory proof-of-concept and into territory where the technology is being stress-tested at genuine wartime scale, with image volumes high enough to expose edge cases and failure modes that small pilot projects might miss.
The sheer volume of imagery also feeds a reinforcing cycle: each confirmed detection trains the underlying model to recognize new variations in mine shape, soil cover, and vegetation interference. That feedback loop is what allows AI systems to accelerate over time in ways that human survey crews, who must visually inspect each square meter of ground, simply cannot replicate. As more sorties are flown and more labels are added, the model can be retrained to handle new seasons, lighting conditions, and sensor payloads, while human deminers use the resulting risk maps to prioritize which fields, roads, and village outskirts should be cleared first.
How Thermal Imaging and Deep Learning Spot Plastic Mines
The technical challenge centers on a specific class of weapon: the PFM-1 “butterfly” mine, a small plastic device scattered by the thousands during conflict. Because it contains almost no metal, it defeats conventional handheld detectors and can blend into leaf litter or soil when viewed from above. Researchers at Binghamton University addressed this problem by flying drones equipped with multispectral and thermal cameras over test sites seeded with PFM-1 mines, then feeding the captured imagery into a Faster R-CNN deep learning model. Their peer-reviewed study in Remote Sensing showed that supervised learning on UAV imagery could automate the detection of scatterable landmines with measurable precision and recall, even when the devices were partially obscured.
The same team emphasized that automated detection was both faster and more reproducible than manual counting from aerial orthoimages, a point they reinforced in a separate university summary of their work. A second Remote Sensing study refined the sensor side of the equation by designing a dedicated thermal imaging protocol for PFM-1 identification. This protocol exploits the fact that plastic mines absorb and release heat differently than the surrounding soil, creating brief but detectable thermal contrast windows around sunrise and sunset. By timing drone flights to those windows, the researchers were able to sharpen the thermal “signature” of each mine, giving the algorithm clearer training examples and reducing false positives from rocks, trash, or vegetation.
Buried Threats and the Push Toward 90 Percent Accuracy
Surface-laid butterfly mines represent only part of the problem. Many explosive devices are buried under sand, soil, or rubble, making them invisible to optical and thermal cameras alike and requiring different sensing physics. To address this gap, a research team mounted ground-penetrating synthetic aperture radar on a compact drone platform. Their technical paper on an airborne GPSAR system for buried targets describes field trials in which the aircraft flew precise, low-altitude tracks and synthesized radar returns into detailed subsurface images. In sandy test environments, the method achieved localization accuracy sufficient to reveal both metallic and non-metallic objects hidden several centimeters below the surface, suggesting a path toward detecting mines that have been covered by shifting soil or debris.
Across multiple research programs, reported detection probabilities in controlled scenarios have clustered around 90 percent. A peer-reviewed literature review of UAV, remote sensing, and AI tools for Ukrainian humanitarian demining found that detection probabilities above 90 percent appeared in several of the studies it surveyed, particularly when sensors and algorithms were carefully tuned to specific mine types and environments. Columbia University scientists working on a parallel initiative that combines drones, geophysics, and artificial intelligence likewise reported a detection rate of about 90 percent in their early field campaigns, according to a project overview from the university. Those figures are not guarantees for every terrain or conflict zone, but they signal that AI-assisted methods are approaching performance levels that demining organizations can meaningfully integrate into their workflows.
What 90 Percent Accuracy Actually Means on the Ground
A 90 percent detection rate sounds impressive in a lab paper, but in a real minefield it means roughly one in ten explosive devices could go undetected. For a farmer returning to a field that once held hundreds of mines, that margin is the difference between safe cultivation and a fatal step. No responsible demining organization treats AI drone surveys as a replacement for follow-up ground clearance; instead, the technology functions as a rapid triage layer that tells human teams where to concentrate their slower, more painstaking work with metal detectors, prodders, and excavation tools. In practice, AI-generated maps can highlight likely hotspots along tree lines, irrigation ditches, and former trench networks, allowing clearance managers to allocate scarce personnel and armored vehicles to the most dangerous plots first.
This triage role also extends to infrastructure and emergency access. Roads, power-line corridors, and approaches to critical facilities can be scanned from the air to identify suspicious patterns that warrant immediate cordoning and manual verification. When an AI model flags a cluster of potential mines near a village, ground teams can be dispatched with a clearer understanding of where to start and which areas might be safe enough for temporary passage. The result is not instant safety but a faster, more informed sequence of risk reduction steps that can shorten the time between the end of active fighting and the cautious resumption of civilian movement, planting, and reconstruction.
From Research Prototypes to Humanitarian Toolkits
Translating these advances from experimental plots and wartime pilots into standard humanitarian practice will require more than clever algorithms. Datasets like Safe Pro’s large-scale image archive and the labeled thermal and radar collections produced by universities must be curated, shared, and updated so that models can be retrained as new mine types, sensor payloads, and environmental conditions emerge. Standards bodies and demining NGOs will need to agree on validation protocols that specify how many ground-truth samples, environmental variations, and negative examples are required before a model can be trusted as a planning aid. In parallel, regulators and donors are likely to ask for transparent reporting on false negatives and false positives, since both have serious implications for safety and cost.
On the operational side, the systems must be packaged for field teams that may lack advanced technical training. That means user interfaces that present risk maps in simple, interpretable layers; workflows that allow deminers to feed back confirmed finds or misses into the training pipeline; and hardware kits that can survive rough transport, dust, and extreme temperatures. Columbia researcher Jasper Baur, profiled in a feature on drone-based demining, has argued that bringing these tools into the field can both accelerate clearance and reduce injuries to human operators. If that promise bears out at scale, AI-guided drones will not eliminate the painstaking craft of demining, but they could reshape it into a more targeted, data-rich discipline, one where every flight and every detection incrementally tilts the balance away from hidden explosives and toward safe, usable land.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
