Your mine detector finds metal, not mines

Between March 1992 and October 1998, the Cambodian Mine Action Centre cleared 89,327 antipersonnel mines, 961 antitank mines and 452,770 items of unexploded ordnance. To find them, deminers dug up 191,737,707 pieces of scrap.

That table sits on page 10 of a RAND report commissioned by the White House Office of Science and Technology Policy, and it is the most clarifying object I know of in this field. Run the arithmetic it invites: for every item a detector called out, there was a 99.7 percent chance it was junk. RAND's accounting of the hours, at three minutes per item to confirm, dig or neutralize, puts 99.6 percent of total clearance time into excavating scrap.

Not finding mines. Digging up bottle caps.

I served as a Counter-IED and Electronic Warfare Officer, so I came to this problem from the countermeasure side rather than the clearance side. What the two share is a habit of mind that took me a while to learn and that almost no coverage of detection technology has: the interesting number is never the detection rate. It is the ratio of things you had to go investigate to things that were real.

What the two sensors physically measure

Start with the tool that is still in every deminer's hands.

A handheld mine detector works by electromagnetic induction. A transmit coil drives a time-varying magnetic field into the ground, that field induces eddy currents in any buried metal, those currents generate a weaker secondary field, and a receive coil picks up the resulting voltage change and turns it into a tone. RAND notes that this uses "the same principles as those first used in World War I and refined during World War II," and that as of its 2003 writing the U.S. military had replaced its standard-issue mine detector only once in forty years.

Read that chain again and notice what is absent. Nothing in it interacts with an explosive. The detector is responding to conductive metal, full stop. RAND states the consequence in a single sentence:

The overwhelming limitation of mine detection using EMI is the inability of EMI systems to discriminate mines from nonmine metal clutter.

And because a modern plastic blast mine may contain only a firing pin's worth of metal, you have to run the detector hot enough to hear that. At which point it also hears, in RAND's list, "shrapnel, bottle caps, bullet casings, and other man-made clutter as well as natural metal in rock."

Ground-penetrating radar was supposed to fix this, and it does fix a different piece of it. GPR does not care about metal. It transmits a radio pulse into the soil and listens for reflections from boundaries where the dielectric constant changes, which is why it can see a plastic-cased mine that an induction coil cannot. In a GPR B-scan a compact buried object appears as a hyperbola, and a great deal of current engineering work, including the automated real-time IED detection system published in Sensors in 2022 that mounted a GPR on a pickup truck and on a maintenance train, amounts to finding that hyperbola reliably.

But look at what the radar is keyed to. It detects a discontinuity in the ground. RAND's table lists the primary source of false alarms for GPR as "natural clutter (roots, rocks, water pockets, etc.)." A tree root is a dielectric discontinuity. So is a stone, a void, a wet pocket, and the backfilled hole left by whoever dug there last.

So the honest summary of the state of the art is this. The metal detector answers "is there metal here." The radar answers "is the ground different here." Neither answers "is there an explosive here," and the gap between those questions is where the years go.

Soil is not a background condition, it is a variable

The second thing coverage flattens is that these are not devices with a performance number. They are devices with a performance number per environment.

RAND cites the International Pilot Project for Technology Cooperation, then the most comprehensive evaluation of commercially available induction detectors. The best performing detector found 91 percent of test mines in clay soil, and 71 percent of the same mines in laterite, the iron-rich soil common across much of the tropics. The worst performing detector found 11 percent in clay and 5 percent in laterite. For reference, the UN standard for mine clearance is 99.6 percent.

Radar has the mirror-image problem, and it is worse in one respect: the direction of the effect is not intuitive. Water attenuates the pulse, so RAND cites work by Koh finding that in wet soils GPR performs poorly for mines buried below roughly 4 cm. Yet theoretical work by Rappaport and colleagues indicated that increased soil moisture can actually strengthen the return from a nonmetallic mine. Both can be true, because what hurts you is not moisture but nonuniform moisture: a wet surface over dry subsurface, or a rough ground boundary. RAND's conclusion is the one to keep:

For the same mine, a given GPR can be very effective or ineffective, depending on soil moisture and mine location.

There is also a design tradeoff with no way out of it. Higher frequency buys resolution and costs penetration. Go too low and small, shallow plastic mines disappear under the surface bounce off the air-to-soil boundary itself.

Why you cannot simply turn the sensitivity down

The obvious response to a 99.7 percent false alarm rate is to make the detector less twitchy. This is the trap, and it is worth stating plainly because it generalizes far beyond landmines.

Detection and false alarms are not independent knobs. They are two coordinates on one curve, and the threshold slides along it. Quieting the detector reduces the number of holes you dig and increases the number of mines you walk over. RAND's judgment on that trade is unambiguous: for humanitarian demining, trading reductions in false alarms for reductions in the likelihood of finding buried mines is unacceptable.

This is the same structure I described in the report on why radar cannot tell a drone from a bird. Detecting a thing and identifying a thing are different measurements with different limits, and a brochure that quotes only the first is telling you the easy half.

What sensor fusion actually bought

The answer the field converged on was to run both sensors together, on the theory that induction and radar have unrelated sources of false alarm, so an object that trips both is likelier to be real.

It works, and the numbers are genuinely good. RAND reports that the Army's Handheld Standoff Mine Detection System, which pairs induction with GPR, achieved in Fort Leonard Wood field testing a probability of detection of 1.00 for metal antitank mines (90 percent confidence interval 1.00 to 0.97) and 0.95 for low-metal antipersonnel mines (interval 0.97 to 0.93), at an average false alarm rate of 0.23 per square meter.

Sit with that last figure, because it is the quietest instructive number in the report. A false alarm rate of 0.23 per square meter is, by my own arithmetic, about 2,300 excavations per hectare cleared. That is the good result, from a fielded dual-sensor system, on a prepared test site, on a day whose soil and weather RAND explicitly warns cannot be extrapolated anywhere else.

Twenty-one years later

In January 2024, a team from Cranfield University's Centre for Defence Chemistry published a practitioner's paper in Heliyon on operational data collection in humanitarian mine action. Their premise, stated in the second paragraph of the introduction, is that "reliable detection capabilities that can discriminate metal, plastic or explosive components from false positives have yet to be developed." Their whole proposal, a standardized clearance data model, is explicitly a workaround:

In the absence of a technological solution to detect and positively discriminate VOEDs from false positive indications, the collection of operational data offers the best prospect for "managing" if not "solving" the problem.

That is 2024 practitioners saying what RAND said in 2003, with two more decades of sensor development behind them.

The same paper carries a field figure I had not seen before, and it is worth the price of admission. Describing one of the only instances where every piece of metal contamination was logged during actual clearance, a trial of the Vallon Minehound dual sensor running from September 2012 to December 2013, the authors report that two detectors checked 197,044 signals, of which 99 percent were clutter, and that for each mine the Minehound found, 32,841 metal signals were investigated. That is real ground, with the good tool.

The 2026 version of the same lesson

Which brings us to the current headlines, which involve drones and machine learning rather than a person with a coil on a stick.

A February 2026 preprint, revised in May and accepted as an oral presentation at IGARSS 2026, benchmarks four classical statistical detectors and a lightweight spectral neural network on drone-borne hyperspectral imagery of inert PFM-1 scatterable mines. The headline result is excellent. The Adaptive Cosine Estimator reaches an area under the ROC curve of 0.989.

Then the authors do the thing that makes this a good paper, and explain why that number is not the number:

However, because target pixels are extremely sparse relative to background, ROC-AUC alone can be misleading; under precision-focused evaluation (PR and AP), the Spectral-NN outperforms classical detectors, achieving the highest AP.

Rare targets, enormous background, and a metric that flatters you until you ask how many of your positives were real. That is the Cambodia table restated in the vocabulary of machine learning. It has a name in this literature, a precision problem under class imbalance, and it is structurally identical to the deminer's problem: an outstanding-sounding score coexisting with an unworkable number of things to go check.

Two caveats on that paper, both of which the authors are clear about and neither of which the trade coverage usually is. It is a preprint of a benchmark study on inert targets. And PFM-1 mines are surface-scattered, so aerial imagery of them is a different sensing problem from finding something buried. Overhead imaging and subsurface detection are not the same capability, and a headline that merges them is doing you no favors.

What I am not claiming

The RAND report is from 2003, and I am not presenting its performance figures as current. Sensors, signal processing and fusion algorithms have all moved since. What I am presenting as current is the structural claim, and for that I lean on the 2024 Cranfield paper, whose authors are practitioners and who state plainly that the discrimination problem is still open.

The Minehound trial figures reach me through that Heliyon paper, which cites a 2014 article by Daniels, Braunstein and Nevard in The Journal of Conventional Weapons Destruction. I could not open the original article: the repository blocks automated retrieval of its PDF. So I report those numbers as the Cranfield authors report them, and flag that I did not read the underlying source myself.

I have not opened the International Pilot Project report or the Koh and Rappaport studies directly either. Those are RAND's citations, quoted as RAND presents them.

Finally, and this should not need saying: nothing here is guidance for handling ordnance. Explosive remnants of war are cleared by trained people with the authority to do it, and the single correct civilian response to a suspected item is distance and a phone call.

The signal

Three things to carry out of this.

First, when you read that something "detects bombs," ask what it physically measures. Metal, a dielectric boundary, a spectral signature in reflected light. If the answer is not "the explosive," the device is inferring, and everything else in the specification is a claim about the quality of that inference.

Second, ask for the false alarm rate in the same breath as the detection rate, and ask what ground it was measured in. A detector without an environment is not a specification. Ninety-one percent in clay and 71 percent in laterite was the same detector.

Third, notice how durable this shape is. The 2003 report, the 2024 field practitioners and the 2026 machine-learning benchmark are describing one problem: a rare target in an enormous cluttered background, where the score that sounds impressive is not the score that governs the work. That pattern runs through threat detection, medical screening, fraud analytics, and every security system that promises to tell you which of tonight's motion events matters. The landmine version is simply the one where the cost of a miss is unambiguous.