UXO (Unexploded ordnance) is among the most enduring and hazardous legacies of armed conflict. Long after fighting ends, buried shells, mines, and munitions remain in the ground and on the seabed, presenting a risk to communities, construction, fishing, and offshore development for decades. Locating these objects safely, before they can cause harm, is one of the most demanding problems in applied geophysics - and one where the cost of a missed detection is measured in lives rather than efficiency.

Magnetic survey is central to how this work is done. Most ordnance contains ferromagnetic material that disturbs the Earth's magnetic field, producing a localised anomaly a magnetometer can measure. In principle, that makes buried ordnance detectable. In practice, turning a field of magnetic readings into a confident, safe decision about what lies beneath is far harder, and the reason returns to a theme that runs through this series: the absence of labelled data.

This post looks at why unexploded ordnance detection is such a difficult setting for conventional machine learning, how a physics-based, unsupervised approach helps, and - just as importantly - where its limits lie in a domain that leaves no room for overstatement.

Why UXO Is the Hardest Case for Labelled Data

Throughout this series we have argued that labelled training data is scarce in geophysical survey. In unexploded ordnance detection, that scarcity is at its most acute, and the reasons are worth stating plainly.

To label a training example, we would need to know with certainty what is buried at a given location - its type, depth, and orientation - before the survey concludes. In ordnance clearance, acquiring that certainty means excavation, and excavation of a suspected munition is precisely the dangerous, expensive step the survey exists to avoid or minimise. The ground truth that supervised learning depends on can, in the most literal sense, only be confirmed by doing the thing the survey is meant to make unnecessary.

There are further complications specific to this domain. Ordnance types vary enormously, from small submunitions to large aerial bombs, each with a different magnetic signature. Burial conditions differ across sites, soils, and seabeds. A model trained on one environment does not transfer reliably to another, and building a broad labelled dataset across the full range of ordnance and conditions is not realistic. In marine settings, where much legacy ordnance now rests, the difficulty and cost of ground-truth verification are greater still.

The result is a domain where the demand for reliable detection is exceptionally high and the supply of labelled data is exceptionally low. This is exactly the gap an unsupervised, physics-based method is designed to address.

What a Physics-Based, Unsupervised Approach Offers

The framework described in the VE3 AI Research paper was not built specifically for ordnance clearance, but its design speaks directly to the constraints of that work. Rather than depending on labelled field examples, it generates synthetic magnetic survey data from the magnetic dipole model, then uses unsupervised clustering to separate anomaly responses from ordinary background.

Three properties of this approach matter particularly in an ordnance context. The first is that it requires no labelled data. Because the synthetic samples are generated from physics, with their properties known by construction, a detector can be developed without waiting on excavation records or controlled test ranges. This directly addresses the domain's central constraint.

The second is interpretability. The features the framework relies on - the maximum, minimum, mean, and peak-to-peak amplitude of the Total Magnetic Intensity signal - are physically meaningful quantities that a geophysicist can reason about directly. When a detection is flagged, its basis can be explained in terms practitioners already trust. In safety-critical work, a result that cannot be interrogated is difficult to act on responsibly; interpretability is not a convenience here but a requirement.

The third is efficiency. The method operates on a compact set of features rather than raw signals and uses clustering that scales predictably as datasets grow. This makes it well suited to a preliminary screening role - reducing a large area of survey data to a shortlist of locations that warrant closer, more expensive investigation. In a field where investigative resources are limited and every flagged location carries real cost, a dependable first pass has clear practical value.

Where the Approach Fits in a Clearance Workflow

It is important to place this kind of method correctly within the wider process of ordnance survey and clearance. It is not a replacement for the established chain of detection, investigation, identification, and disposal carried out by trained specialists. It is a tool that can strengthen the earliest analytical stage of that chain.

A typical magnetic survey produces a large number of candidate anomalies, only some of which correspond to genuine items of interest. Many are geological features, background variation, or surface debris. The task of separating the anomalies that merit investigation from those that do not is where an unsupervised, physics-based first pass can contribute. By grouping coherent anomaly responses and setting genuinely ambiguous points aside as noise, the method can help prioritise which locations deserve the attention of more sensitive equipment and expert interpretation.

Understood this way, the contribution is not autonomy but triage. The method helps direct limited, high-value resources toward the locations where they are most needed, while leaving the decisions that carry safety consequences with the specialists trained to make them. That framing is not a limitation to apologise for - in this domain, it is the responsible way for any automated tool to operate.

The Limits That Must Be Stated Clearly

In most applications, discussing the boundaries of a method is good practice. In ordnance detection, it is an obligation, because the consequences of overstating a capability are severe. Several limits deserve explicit acknowledgement.

• Validation on synthetic data.  The framework has been demonstrated on physics-based synthetic data. Real ordnance sites introduce complexities - heterogeneous geology, variable soil magnetisation, corroded or fragmented objects, and interference from surrounding metal - that a controlled simulation cannot fully reproduce. Performance on synthetic data indicates promise, not field-readiness.

• The dipole approximation.  Representing an object as a single magnetic dipole is a reasonable first approximation, but real ordnance with complex geometry can produce signatures that depart from the ideal, particularly at close range. For preliminary screening this is generally acceptable; for precise identification it is a known simplification.

• Preliminary, not definitive.  The method supports early anomaly identification. It does not classify ordnance type, confirm the presence of a live item, or make any determination that bears directly on safety. Those judgements remain firmly with qualified personnel and appropriate equipment.

None of these caveats diminishes the value of the approach in the role for which it is suited. They define that role precisely. A method that helps prioritise where scarce investigative effort should be directed, without any labelled data, and with fully interpretable outputs, is a genuine contribution - provided its outputs are always treated as a first analytical step within a process that specialists control from end to end.

Beyond Ordnance: The Same Pattern, Different Stakes

Although unexploded ordnance is the most safety-critical application, the same underlying pattern - scarce labelled data, irregular anomaly signatures, and a need to separate signal from background - appears across several fields of subsurface survey. The value of a physics-based, unsupervised method is that the same principles transfer, even as the specific stakes change.

In marine survey, the approach applies to mapping submerged metallic debris, wreckage, and seabed infrastructure, where ground-truth verification is especially costly. In infrastructure inspection, it supports locating buried pipelines, cables, and structural elements whose precise position or condition is uncertain. In mineral exploration, it can help screen for ferromagnetic geological targets across large survey areas. In each case, the recurring advantage is the same: a scalable, interpretable first pass that does not depend on labelled data that may not exist.

What unites these applications is not the specific objects being sought but the shape of the problem. Wherever we must find something buried, without the labelled examples that conventional machine learning assumes, and where the cost of investigation makes prioritisation valuable, a physics-based unsupervised framework offers a principled starting point.

A Foundation, Used Responsibly

Unexploded ordnance (UXO) detection sets a high bar. It demands reliability in an environment that withholds the labelled data reliable models are usually built from, and it permits no exaggeration of what a method can do. Judged against that bar, a physics-based, unsupervised approach is valuable precisely because it is honest about its scope: it generates its own training data from established physics, produces outputs a specialist can interpret, and takes on the preliminary screening role rather than claiming the decisions that belong to trained personnel.

Used in that spirit, the framework is not a solution imposed on a difficult problem, but a foundation offered to the people who work on it - a way to direct scarce, high-value effort more effectively, grounded in physics that the field has trusted for generations. In a domain where getting it wrong carries the gravest consequences, that combination of usefulness and restraint is exactly what a responsible tool should offer.

Read the full methodology - including the complete DBSCAN parameter analysis and PCA cluster visualisations, in the research paper: Data-Driven Buried Anomaly Detection Without Annotated Samples.