Over-the-horizon USV bathymetry: autonomy, data quality and QC

What NAVOCEANO actually bought

Saildrone has been awarded a US$10 million task order by NAVOCEANO, the US Navy’s Naval Oceanographic Office, to collect bathymetric survey data in international waters of the Atlantic. The work uses the Saildrone Surveyor, a 20-metre (65-foot) uncrewed surface vehicle carrying a deep-ocean multibeam sonar, with operations beginning in 2026 and running through most of 2027. The company has been delivering bathymetric surveys to US government customers since 2019.

The framing is familiar to anyone tracking Seabed 2030: the source notes that only 28.7% of the world’s oceans are mapped to modern standards, and Saildrone’s Senior Director Ocean Mapping, Kitch Kennedy (USN Capt, retired), points to vast Atlantic regions where “traditional survey deployments are long and the conditions are unpredictable” as precisely the mission the platform was built for. That is the sales argument. The engineering question for any survey buyer is narrower and more useful: when does an over-the-horizon USV actually hold data quality, and how do you prove it did.

Why the data-quality exposure moves to the buyer

With a crewed vessel, the surveyor-in-charge is standing next to the acquisition system. Sound velocity casts get taken when the water column changes, a bad patch test gets re-run the same watch, and a systematic bias shows up on the online helmsman display before it contaminates a week of lines. The quality-control loop is tight because the people who can act are physically present.

Over the horizon, that loop stretches. Satellite bandwidth will not carry a full deep-ocean multibeam data rate, so what comes ashore in near real time is heavily subsampled – coverage tracks, health telemetry, perhaps decimated soundings. The full dataset is reviewed on data offload, which may be weeks or months after acquisition. The consequence is structural: a USV can run clean lines for a long time while an undetected refraction or motion problem quietly degrades every one of them, and nobody finds out until the data is in the office.

That shifts the burden of assurance onto the buyer’s specification. The platform vendor can demonstrate the vehicle performs. Whether the delivered data meets a stated IHO S-44 order, with a defensible uncertainty budget and complete coverage, is a contractual and QC question that the client has to define up front – because there is no surveyor on deck to catch the gap for you.

What really drives the autonomy-versus-crewed choice

The decision is not ideological. It turns on four engineering realities, and they do not all point the same way.

Depth sets the error budget, and deep water is forgiving on the terms autonomy is weak at. S-44 Edition 6.0 expresses total vertical uncertainty as the quadrature sum of a depth-independent term and a depth-proportional term. For Order 2 reconnaissance work – the realistic target for open-ocean transit-corridor mapping – the depth-proportional term dominates completely in several thousand metres of water. The allowable vertical uncertainty at abyssal depth runs from tens of metres to over a hundred metres. This matters because the weakest part of an uncrewed workflow, the vertical reference, is cheap to satisfy here. Without a tide gauge in international waters you work ellipsoidally, with GNSS precise point positioning and a geoid or ocean-model separation surface. PPP delivers vertical accuracy at the decimetre level after convergence, which is negligible against a multi-metre allowable. The datum and separation work that eats budget on shelf and nearshore programmes – the kind of effort we have set out in the geodetic cost hidden inside harmonised megasurveys – is far less punishing in deep water.

Sound velocity is where autonomy actually bites. In deep-ocean multibeam, refraction through an imperfectly known water column is the dominant systematic error on the outer beams. A crewed vessel manages this with regular CTD or SVP casts. A USV typically cannot deploy a full-depth cast underway without a winch and crew; it relies on a hull-mounted surface sound-speed probe to steer the array, plus a modelled profile from oceanographic climatology or an operational ocean model to correct the full raypath. Surface sound speed fixes beam forming at the transducer face. It does nothing for refraction deeper in the column. In a dynamic frontal region the residual refraction error can smile or frown the swath and pull outer-beam soundings out of tolerance even while the inner beams look clean. This is the single biggest technical risk to absolute accuracy in an uncrewed deep-water survey, and it is invisible in coverage telemetry.

Coverage rate favours the platform more than its speed suggests. A slow, largely wind-driven platform runs at a few knots where a crewed launch surveys at eight to ten. But swath width scales with depth – roughly four to six times water depth for usable full coverage – so in thousands of metres the swath is kilometres wide, and the vehicle acquires continuously for months with no crew change, no weather return to port, and no demobilisation clock. Over a long campaign the area-per-euro can beat a crewed vessel even at a third of the speed.

Latency tolerance is the hinge of the whole decision. If the deliverable supports a planning cycle that can wait for data offload, autonomy is attractive. If anyone downstream needs near-real-time charted depths – route clearance ahead of an asset move, an emergency response – the QC latency of an over-the-horizon platform is a liability, and no amount of endurance compensates.

Where buyers get over-the-horizon surveys wrong

1. Equating swath coverage with data quality

Coverage tracks are the one product that transmits cheaply over satcom, so they become the thing everyone watches. Full bottom coverage with zero holidays tells you the seabed was ensonified. It tells you nothing about whether the soundings meet uncertainty. A refraction bias, an uncorrected latency, or a degraded motion solution produces complete, confident, and wrong coverage. Treat the coverage map as a necessary condition, never a sufficient one.

2. Assuming sound velocity is handled because there is a probe on the hull

A surface sound-speed sensor is standard and necessary, and its presence reassures the untrained eye. The question that matters is how the full-column profile is derived, how often it is validated against an actual cast, and what refraction residual that implies on the outer beams in the survey area’s oceanography. If the answer is “modelled climatology, never validated,” the outer thirty per cent of every swath is carrying an unquantified systematic error.

3. Trusting line keeping and coverage against a plan drawn in still water

Survey line plans assume the platform holds the line and the speed. A low-speed autonomous vehicle in a strong boundary current or persistent cross-wind can be set off track, lose speed over ground, and open gaps between adjacent swaths that no one sees until processing. Planned line spacing is an intention, not an outcome, and the achieved overlap is what determines whether you have genuine coverage.

4. Letting calibration validity lapse over a long deployment

A patch test – roll, pitch, yaw and latency – is the foundation of multibeam accuracy, and on a crewed job it gets re-run when something changes. You cannot send a technician out to re-run one on a vehicle three weeks into an ocean crossing. That places the whole weight of geometric integrity on a stable, well-characterised mounting, an aided inertial solution that does not drift, and a monitoring regime built from crosslines and overlap comparisons rather than fresh calibration.

5. Underestimating who has to carry QC authority, and when

On a crewed vessel the client representative and the party chief hold QC authority in real time. In an uncrewed, delayed-data model, that authority has to be redefined – who reviews the automated onboard flags, who decides a line needs reflying, and how a reflight even happens when the asset is autonomous and far away. Buyers who do not assign this end up accepting data after the fact with no leverage to correct it.

How to specify and accept an over-the-horizon USV survey

The platform is not the variable you should be managing. The specification is. Put the following in the scope of work and the acceptance criteria, in concrete terms.

• State the target S-44 order and the full uncertainty model, not just the order name. Require a total propagated uncertainty (TPU) budget that explicitly includes the sound-velocity contribution and the motion-solution contribution, and require delivery of a CUBE or CHRT uncertainty surface, not just a depth grid. An uncertainty surface that ignores refraction is a decoration.

• Mandate a documented sound-velocity strategy with a residual bound. Require the profile source (surface probe plus named ocean model or climatology), a validation cadence against actual casts where the vehicle can take them, and a stated worst-case refraction residual on the outer beams for the survey area. If the vehicle cannot cast, the model uncertainty has to be carried in the TPU and the outer-beam swath may need to be trimmed in processing.

• Specify the vertical-referencing workflow by name. Ellipsoidally referenced survey with PPP, a named geoid or separation model, stated PPP convergence handling, and QC on the GNSS solution. In deep water this is the easy part of the budget – make sure it is documented rather than assumed.

• Require coverage and holiday reporting against plan, with a maximum allowable gap. Define the achieved-overlap threshold and the holiday size that triggers a reflight, and require the achieved line geometry, not the planned one, in every progress report.

• Set a crossline QC regime and a bias-monitoring method. Require crosslines for independent depth comparison, report the crossline statistics against the S-44 allowable, and require that systematic trends in overlap comparisons are flagged and reported during the campaign, not discovered at offload.

• Fix the QC authority and the data-latency checkpoints contractually. Name who reviews onboard automated QC flags, define the subsampled-telemetry checkpoints at which the client can halt or redirect, and set the data-offload and chain-of-custody arrangements for data recovered months after acquisition.

• Let latency tolerance veto the method. If the deliverable feeds a decision that cannot wait for offload, do not choose an over-the-horizon platform with no real-time QC path, regardless of its endurance economics.

Autonomy is changing who can realistically bid ocean-scale bathymetry, and it rewards operators who can hold a fleet at sea for a year – a shift that feeds directly into the long-running question of whether scale decides the offshore survey market. For the buyer, though, the call is simpler and older than the technology. Specify the data quality you need, specify how you will prove it was met, and then the platform – crewed or not – becomes an implementation detail rather than a leap of faith.

---

Based on: Saildrone secures US Navy deal for Atlantic seabed mapping