The ground decides what your mine detector finds

A detector datasheet gives you a number: detection depth, sensitivity, probability of detection. What it almost never gives you is the soil that number was measured in, and without that, the number is closer to a mood than a specification.

I served as a Counter-IED and Electronic Warfare Officer, which is the countermeasure side of this problem rather than the clearance side. The habit the two share is asking what the environment contributes to a sensor's output before asking what the target contributes. In Report 097 I argued that a mine detector finds metal and not mines, and that the false alarm rate is the number that governs the work. This report is the sentence underneath that one: the false alarm rate is largely a property of the ground, not of the device.

The trial that separates the device from the dirt

The cleanest demonstration I know of comes from a trial run by BAM, Germany's Federal Institute for Materials Research and Testing, at the Croatian Mine Action Centre's test site in Benkovačko Selo near Benkovac, in May 2005. It was conducted under CWA 14747:2003, the CEN Workshop Agreement that specifies how metal detectors for humanitarian demining are to be tested.

The design is the interesting part. Four detector models from four manufacturers (CEIA, Ebinger, Foerster and Vallon, anonymized in the write-up as U, X, Y and Z) were run by four experienced CROMAC deminers across four lanes of 1 m by 29 m. Lanes 1 and 2 held highly magnetic soil from the area around Obrovac. Lanes 3 and 4 held magnetically neutral soil from around Sisak. Targets were deactivated PMA-2 and PMA-1A mines buried in fixed depth tiers. The PMA-2 is described in the paper as a minimum-metal mine, and the hardest to detect in south-eastern Europe.

The blind reliability tests used a fractional factorial design based on a Graeco-Latin square, which is a formal way of saying that detector, operator, lane and run order were arranged so their effects could be pulled apart afterward instead of being confounded with each other. That is the whole reason the trial can answer this question at all.

Two results fall out of it. The first:

It can be seen that the false alarm rate increases in soils with stronger magnetic properties, that is, in lanes 1 and 2 (soil from Obrovac). It is likely that the probability of detection is reduced as a consequence of ground compensation.

The second is the one that makes the first land:

The diagram on Figure 4 clearly shows that all operators performed similarly: no significant differences were detected between the deminers' results.

Four different humans, no measurable spread. Four lanes of dirt, a visible one. The soil was the loudest variable in the room.

What magnetic soil actually does to an induction detector

A handheld mine detector drives a time-varying magnetic field into the ground and listens for the secondary field that buried metal returns. Nothing in that chain is specific to a mine, and nothing in it is specific to metal either. It is specific to anything that responds to a changing magnetic field.

Soil does. Iron-bearing minerals give soil a magnetic susceptibility, and in strongly weathered tropical and subtropical soils that susceptibility is both large and frequency dependent, meaning the soil's response varies with the driving frequency in a way that can look like a real target. The BAM authors state the consequence plainly:

The main source of false alarms are the soils with frequency dependent magnetic susceptibility, causing alarms without presence of a metal fragment.

Note the phrase "without presence of a metal fragment." This is not the scrap-clutter problem from Report 097, where the detector was right about the metal and wrong about the mine. This is the detector alarming on nothing at all, on the dirt itself.

Kazunori Takahashi, Holger Preetz and Jan Igel, working at Germany's Leibniz Institute for Applied Geophysics, ran a four-soil comparison (laterite, an engineered magnetic sand, and two humus soils) and reached the same conclusion from the materials side. Their summary of it is compact:

Magnetic susceptibility, particularly its frequency dependence, is the most influential property on metal detectors.

Ground compensation is a trade, not a fix

Every serious detector has a ground compensation or ground balance function, and every operator uses it. It is also the mechanism by which the soil quietly takes your detection depth. The BAM description is worth reading closely:

Most metal detectors today have some ground compensating abilities, which means that they can decrease their sensitivity to soils with much smaller decrease of sensitivity to metal.

"Much smaller decrease" is not "no decrease." Compensation buys down the false alarms by turning down the gain on a response that the soil and the target partly share. You cannot subtract the ground for free. That is why the Croatia result reads the way it does: false alarms up in the magnetic lanes, and detection probability likely down as the price of holding those false alarms in check.

It is also why "detects to 20 cm" is a claim about a soil, and why a detector that grades well on a test range in loess can grade badly on laterite without anything about the device having changed. Report 097 cited International Pilot Project figures in which the best detector managed 91 percent in one ground and 71 percent in another. Same hardware, twenty points.

GPR fails differently, in different dirt

Ground-penetrating radar is often proposed as the answer to magnetic soil, and it is, in the narrow sense that GPR does not care about magnetic susceptibility. It cares about the dielectric permittivity of the ground, and specifically about how much that permittivity varies from point to point, because a buried mine is detected as a contrast against its surroundings. Ground full of contrasts is ground full of decoys.

In the four-soil comparison, the soil expected to be hardest for GPR was not the laterite that wrecked the metal detectors. It was a loamy humus forest soil with high stone content, precisely because of the large spatial variation in its dielectric permittivity. The authors' summary of the GPR pattern:

The FAR reduction (positive feature) is almost constant over all the soils. However, the POD loss (negative feature) increases with soil difficulty.

So the two sensors do not merely degrade at different rates. They degrade in different soils, and they degrade in different directions: the metal detector's magnetic-soil failure shows up mostly as false alarms, the radar's inhomogeneous-soil failure shows up mostly as misses. A dual-sensor system is not immune to either. It inherits both failure modes, and the ground has to be characterized for both.

There is a standard for the dirt

Here is the detail that convinced me this deserved its own report. The demining community did not treat soil as a footnote to detector testing. It gave soil its own standard.

CWA 14747-1:2003, the metal detector test and evaluation protocol, already says the quiet part in section 8.1.1:

The soil in which the mines are buried will in general hamper the detection performance of the detector.

And it does not leave the soil undescribed. Section 8.1.4 requires that magnetic susceptibility and electrical conductivity be measured in situ with calibrated instruments across the test area, together with moisture content in the top 0.2 m, reported in specified units. Then it adds a requirement most people would never think to write down:

The measurements shall be repeated after any occurrence that may change the soil properties, e.g. rain.

Rain changes the answer. Not the detector's answer about the mine, the test's answer about the detector.

Five years later that concern became a document of its own: CWA 14747-2:2008, "Humanitarian mine action, Test and evaluation, Part 2: Soil characterization for metal detector and ground penetrating radar performance." Its stated purpose is to give mine action programmes simple procedures for assessing how soils affect metal detectors and dual sensors, to help them recognize problematic ground, and to give test designers a compilation of the relevant soil properties, methods for determining them, and classifications based on their effect on detector performance.

Read that as a statement about evidence rather than about dirt. An entire part of an international test standard exists because a detector result is not interpretable without a soil result attached to it.

One more thing the trial found

A side finding from Benkovac deserves a paragraph, because it cuts against a reflex most of us have. Almost every metal detector trial before 2005 measured maximum detection height once, without repetition. BAM repeated it, and the spread was not small. Their diagnosis:

Further investigations have shown that the variability of the measurements is not caused only by the operators (deminers) nor by their personal differences, but mostly by the instability of the hardware of the devices.

The instrument itself does not return the same number twice. A single unreplicated detection-depth measurement, the kind that fills marketing tables, is one draw from a distribution nobody showed you.

The signal

Three things to carry out of this.

First, a detection specification with no soil attached is incomplete in a way that cannot be repaired later. Ask what ground the number came from, and ask whether that ground was measured or merely described. "Sandy loam" is a description. A susceptibility value with its frequency dependence is a measurement.

Second, when a sensor has an automatic environment-compensation feature, ask what it costs. Ground balance, clutter rejection, adaptive thresholds and background subtraction all work the same way: they suppress a response that the environment and the target partly share, and the target pays part of the bill. A system that reports only its post-compensation performance has hidden the trade, not avoided it.

Third, notice that the demining field solved the reporting problem the hard way, by standardizing the environment measurement and publishing it as a separate part of the standard. Very few sensing fields have done that. Threat detection, environmental monitoring, medical screening and the whole category of security analytics all quote performance figures whose environment is unstated and unmeasured. When you see a detection rate with no environment behind it, you are not looking at a weaker specification. You are looking at a different kind of claim.