At least **57 countries** still have landmines buried in their soil. In 2024, those mines killed 1,945 people and injured 4,325 more. Nearly 90% were civilians. Nearly half were children.
For decades, the work of finding and clearing them has relied on methods that are slow, expensive, and dangerous: metal detectors that struggle in mineral-rich soil, ground-penetrating radar that fails in wet or uneven terrain, trained dogs, and manual probing — all of which require human beings to work, systematically, through landscapes that can kill them.
Researchers at **Rochester Institute of Technology (RIT)**, in collaboration with the nonprofit **Demining Research Community**, have built something that could fundamentally change that equation.
**What They Built**
The core innovation isn't a single device. It's a dataset — the first **publicly available benchmark dataset** specifically designed to train AI algorithms to detect landmines from drone imagery.
The dataset was built with painstaking care. Researchers flew drone surveys over test fields using a comprehensive array of sensors:
- 📷 **RGB cameras** (standard visual imagery) - 🌡️ **Thermal imagers** (detecting heat signatures) - 🌈 **Multispectral and hyperspectral sensors** (capturing invisible light bands) - 📡 **LiDAR** (laser-based 3D mapping) - ⚡ **Electromagnetic detectors** (metal detection from the air)
Crucially, every data point in the dataset is tagged with **precise ground truth information** — the exact known location of inert (safe, non-explosive) mines placed for testing purposes. This means AI models trained on the dataset can be rigorously evaluated: not just "does the algorithm detect something?" but "is it detecting the mine, and only the mine?"
Additional datasets were collected at the **Royal Military Academy of Belgium**, using replicas of PFM-1 butterfly mines — the Soviet-designed scatter mines infamous for maiming civilians in Afghanistan, Angola, and Cambodia — to ensure the AI learns to recognise some of the world's most common, and most dangerous, mine types.
**The AI Twist: Knowing What It Doesn't Know**
One of the most sophisticated elements of the research, led by **Sagar Lekhak** (a PhD student in RIT's Imaging Science Department) and **Professor Emmett Ientilucci**, is the focus on *uncertainty quantification*.
AI detection models often produce confident-sounding outputs even when they're uncertain. In mine detection, a false negative — "nothing there" when there is — can kill a deminer. The RIT team's algorithms are specifically designed to flag the **confidence level** of each detection: not just "mine found" but "mine found, 94% confidence" versus "possible anomaly, 61% confidence."
This gives human deminers far better information about where to send people and where to rely on machine surveys alone — dramatically reducing unnecessary risk.
**Why the Dataset Matters More Than the Technology**
The drone sensors and AI algorithms aren't entirely new. What was missing — what has been missing for years — was training data good enough to make them work reliably.
By making the benchmark dataset **publicly available**, the RIT team has opened the field. Any researcher, any humanitarian organisation, any government demining program can now use this data to train, test, and improve their own AI detection systems. The improvement compounds: better data produces better models; better models in more hands means faster progress worldwide.
"Turning it from a slow, dangerous practice into a safer, smarter, and more scalable process," is how Lekhak describes the goal — one that can *"turn post-conflict landscapes back into places where life can grow again."*
**The Stakes**
Croatia spent 30 years and €1.2 billion clearing its landmines. Bosnia-Herzegovina still has an estimated 100,000 mines in the ground, 30 years after its war ended. Ukraine is now accumulating a landmine burden that dwarfs either — with an estimated 30% of the country's territory affected.
The manual clearance rate — with existing tools — cannot match the pace at which mines are being laid or the backlog already accumulated. Drone-based AI detection won't replace deminers. But it can tell them, far more quickly and safely, exactly where to go.
Somewhere in a test field in New York State, drone sensors are already learning to read the ground. The fields they'll eventually help clear are in 57 countries, waiting for this to be ready. 🕊️
*Sources: National Today / RIT Today (nationaltoday.com, March 6, 2026) · DroneXL · The Conversation (theconversation.com) · Good Good Good (Week of March 14, 2026) · Demining Research Community · Sagar Lekhak, Emmett Ientilucci, Rochester Institute of Technology*
