AZARELBefore a cable is laid or a wind farm foundation is driven, somebody has to survey the seabed for unexploded ordnance. Large areas of the North Sea and Baltic contain munitions dumped after two world wars, and the survey is not optional. Militaries face the same problem in a more urgent form: finding objects deliberately placed to be hard to find.
Both are search problems with a brutal prior. The area is enormous, the target is small, and the number of targets is very close to zero. That combination means the performance metric that decides whether a system is usable is not detection rate. It is false alarm rate. A classifier that finds every mine and flags four hundred rocks per square kilometre has produced work, not answers.
The shadow does the work
Side-scan and synthetic aperture sonar produce an image where a proud object shows as a bright return followed by an acoustic shadow — the region the sound could not reach because the object blocked it. For classification, the shadow is usually the more informative half. The bright return tells you something is there and reflects; the shadow encodes height and shape, and it is far less sensitive to the object’s surface than the return is.
This has a consequence people outside the field find surprising: the geometry of the survey changes the classifier’s input. Shadow length depends on altitude and grazing angle. Fly the same object at a different height and you have produced a different training example. Vehicle control and perception are not separable problems here, which is exactly the kind of coupling a simulator is good at exposing and a static dataset is bad at representing.
Why the data does not exist
You cannot buy a large, well-labelled dataset of mine-like objects on varied seabeds. Where it exists it is classified. Where it is public it is small. Collecting your own means a vessel, a survey, ground truth by other means, and a set of objects you are allowed to place on the seabed — a programme, not a dataset download.
So the field is in the position of trying to apply modern detection methods with a fraction of the data those methods assume. Synthetic data is not a nice-to-have here. It is the only route to volume that exists.
What the model has to get right to be worth anything
A synthetic sonar image that looks clean will train a classifier that fails immediately on real data. The artefacts are the content. That means modelling, deliberately:
- Side lobes at realistic levels, because they produce the false returns that a real operator learns to discount.
- Range ambiguity, where a return from the previous ping arrives inside the current one.
- Speckle as coherent interference rather than as added noise, because it has structure and noise does not.
- The shadow as its own labelled class, not as background. If the shadow is the signal, the annotation has to say so.
- Seabed variation — sand ripples, rock, mud — because clutter is what the false alarm rate is made of.
Coverage is half the problem
Detection gets the attention; coverage decides the outcome. A survey that leaves gaps between adjacent lines has not searched the area, and whether it leaves gaps depends on how well the vehicle held its line in current, how altitude varied over changing bathymetry, and whether the planner accounted for the swath narrowing directly beneath the vehicle. These are control and planning questions with a measurable answer, and they belong in the same scored task as the detection.
What this does for the field
Detection results in this domain are notoriously hard to compare. Different seabeds, different sonars, different altitudes, different object sets, and results reported on data nobody else can obtain. A frozen, versioned synthetic benchmark does not replace trials at sea, but it gives two groups a way to disagree productively — to run the same scenario, with the same objects, at the same altitude, and see whose approach holds up.
That is a modest-sounding contribution. In a field where most published numbers cannot be reproduced by anyone outside the group that produced them, it is not.
