In conflict zones and in regions that have seen prior warfare, small plastic objects shaped like maple seeds can remain active for years, hidden in grass, snow, or soil. These PFM-1 antipersonnel landmines, often called butterfly mines, are scatterable devices designed for wide-area dispersal rather than precise placement. Their plastic construction and surface or near-surface positioning makes them especially difficult for conventional demining tools to locate.
Antipersonnel landmines are small explosive devices intended to injure or kill people on foot. A typical example contains a main charge triggered by pressure on a fuse or the mine body itself. The goal on the battlefield is often not to kill but to wound: an injured soldier requires more resources for evacuation, medical care, and ongoing support than a fatality, thereby straining an opponent’s logistics and morale. Scatterable mines amplify this effect by allowing rapid deployment over large zones. Artillery shells, rockets, aircraft dispensers, or even drones release clusters of mines that spread according to ballistic paths or simple dispersal mechanisms. The PFM-1 exploits this approach through its distinctive wings. When released from height, the wings cause the mine to autorotate and glide somewhat like a maple seed, distributing devices unpredictably across fields, roadsides, or villages rather than in orderly rows. Once on the ground, the mine sits exposed or only lightly covered, ready to detonate under foot pressure.
The evolution toward nonmetallic construction made these weapons far harder to find. During World War II, German forces introduced the Schu-mine 42, a wooden-cased blast mine whose lack of metal defeated the magnetic detectors then in use by Allied troops. Later variants incorporated more glass or plastic elements. By the late twentieth century, the Soviet PFM-1 represented a further step: an almost entirely plastic body with minimal or no metal components, paired with the aerodynamic shape suited to mass aerial scattering. Plastic casings resist standard metal detectors entirely. Geophysical methods such as electromagnetic induction, ground-penetrating radar, and magnetometry, which work well on metallic objects by sensing conductivity or magnetic anomalies, perform poorly on low-metal or fully plastic mines, especially when the devices sit on the surface or among heterogeneous soil and vegetation. The result is slower, more dangerous, and more expensive clearance work in post-conflict settings where resources are already limited.
Remote sensing from uncrewed aerial vehicles equipped with ordinary RGB cameras provides one avenue for addressing surface-laid scatterable mines. High-resolution photographs can capture the distinctive shape and color contrast of a PFM-1 against its surroundings, provided the drone flies low enough, typically 10 to 20 meters above ground, for the sensor to resolve the roughly 12-centimeter-wide object. Manual review of thousands of such images is impractical. Machine-learning object detection offers a way forward. Algorithms learn visual patterns from labeled examples and then scan new images for similar features. One efficient architecture, known as You Only Look Once or YOLO, processes an entire image in a single forward pass through a convolutional neural network. It predicts both the class of an object and its bounding-box location simultaneously, making it fast enough for near-real-time use on modest hardware.
Sharifa Karwandyar, Thomas J. Pingel, and Alex Nikulin developed and tested exactly this kind of system in their paper titled “Deep Learning and Multiview-Based Detection of Scatterable PFM-1 Landmines: Performance, Out-of-Sample Evaluation, and Field Readiness,” published in the journal Geomatics in 2026. They collected more than 5,800 high-resolution RGB images of an inert PFM-1 mine (marked with a Cyrillic “Y” on one wing for training identification) placed across varied campus terrains including bare ground, gravel, short grass, light snow, and plant litter. Data came from multiple sensors, among them consumer drones and smartphones, under different lighting and seasonal conditions. Images were labeled with bounding boxes around the mine. Two YOLOv11 models were trained on individual frames rather than stitched Ortho mosaics, preserving fine detail that can blur during three-dimensional reconstruction. One model learned only the PFM-1 class plus background scenes. The second incorporated the same mine images plus the large public COCO dataset of everyday objects, giving the network practice distinguishing the mine from common environmental clutter.
To test realism, the team gathered entirely new out-of-sample imagery from fresh locations and sensors never seen during training. This set included both the original inert mine and low-cost 3D-printed replicas. The replicas were created from a publicly shared design, validated against a three-dimensional scan of the inert mine, painted to match surface appearance, and produced for roughly one dollar each in filament. Such replicas enable safe, scalable data collection without handling live ordnance. Detections from geotagged photographs were then mapped. Simple projection using camera altitude, yaw, and field-of-view data located each detection on the ground with mean error around 1.75 meters when metadata were used, sufficient to highlight clusters for follow-up inspection. Kernel-density mapping further emphasizes likely concentrations of mines.
Performance on the internal validation and test splits reached high levels, with precision and recall often in the 76 to 94% range depending on the model and split. On the truly unseen out-of-sample images, however, recall fell sharply to between 14 and 24% while precision remained stronger, between 74 and 80%. The model trained with the broader COCO dataset performed modestly better on the unseen data than the specialized model. Three-dimensional replicas were detected at least as well as the inert mine in the out-of-sample tests. These outcomes underscore that models can appear highly capable when evaluated only on data drawn from the same collection campaign yet degrade noticeably when confronted with new backgrounds, lighting, sensors, or slight variations in mine appearance such as paint texture or the absence of training-specific marking.
The workflow carries a practical advantage for field use. Training requires hours to a day on capable graphics hardware, but once complete the model runs locally on a consumer-grade laptop or edge device. No internet connection or cloud upload is required after training. Photographs can be processed while the drone is still airborne or immediately afterward, allowing operators to mark potential locations on the spot. This capability aligns directly with the non-technical survey phase of humanitarian mine-action standards, which aims to identify suspected hazardous areas through desk study, community interviews, and initial field assessment without intrusive clearance. A rapid aerial pass that flags clusters of detections can reduce the size of areas requiring detailed technical survey and manual clearance, thereby conserving limited resources and lowering risk to personnel.
Because scatterable mines like PFM-1 are often deployed in disorganized patterns over wide areas, even partial detection of surface examples can indicate nearby hazards and guide prioritization. The approach does not replace trained deminers or established geophysical and manual methods; rather, it supplies an accessible, low-cost layer of information that integrates with local knowledge of terrain and historical records. The use of individual Multiview images and modest computational needs further lowers barriers in regions where full three-dimensional reconstruction software or high-end servers may be unavailable.
Real-world translation will require additional validation. Performance was measured under relatively favorable conditions on a university campus. Denser vegetation, urban rubble, heavy occlusion, or partial burial by soil or snow can reduce the visible target area and lower detection rates. Future refinements could include targeted data augmentation, synthetic imagery to simulate varied textures and partial concealment, and expanded testing across additional mine types. Controlled comparisons of model components and field trials conducted in partnership with operational demining organizations would strengthen confidence before wider adoption.
The study demonstrates a replicable pathway: collecting diverse imagery with accessible drones and cameras, training on single frames, validate rigorously on unseen data, map results using standard image metadata, and deploy locally without connectivity. By showing both the promise and the current limits of this optical deep-learning method for surface-laid scatterable mines, the work contributes a concrete benchmark for evaluating similar systems in humanitarian contexts. Continued development along these lines could help accelerate the initial narrowing of suspected hazardous areas, making subsequent clearance safer and more efficient where conventional tools alone fall short.
This article is shared at no charge for educational and informational purposes only.
Red Sky Alliance is a Cyber Threat Analysis and Intelligence Service organization. We provide indicators of compromise information (CTI) via a notification/Tier I analysis service (RedXray) or an analysis service (CTAC). For questions, comments or assistance, please contact the office directly at 1-844-492-7225, or feedback@redskyalliance.com
Weekly Cyber Intelligence Briefings:
- Reporting: https://www.redskyalliance.org/
- Website: https://www.redskyalliance.com/
- LinkedIn: https://www.linkedin.com/company/64265941
Weekly Cyber Intelligence Briefings:
REDSHORTS - Weekly Cyber Intelligence Briefings
https://attendee.gotowebinar.com/register/7855487668891299929
Comments