Validating a seabed mobility atlas without ground truth

Why direct validation is impossible

The quantity the atlas produces — the fraction of time the seabed is set in motion, cell by cell, over more than thirty years — is observable nowhere. Nobody deploys thirty years of instruments across 2,581 cells. The temptation is then to validate nothing and present the map as it stands, which is what a large share of screening studies do.

The way out is to break the chain apart and attack each link with a source that was not used to build it. Three links are open to attack: the forcing that enters the computation, the observable morphological consequence that should follow from it, and the weakest assumption in the model. None of these three confrontations proves the atlas. Together, they bound what it is worth.

Confrontation 1 — the forcing, against the SHOM tidal current atlas

Bed shear stress depends on the current. If the current is wrong, everything is. So we compare the model's hourly currents to the 51 points of the SHOM "Tidal currents of the French coasts" product (C2D, 2005 edition, Open Licence) that fall inside the tile — a product that enters the pipeline at no point. Spring tides (coefficient 95) and neap tides (coefficient 45) are reconstructed from the multi-year hourly cache, then composited hour by hour from −6 h to +6 h around high water at the reference port (Calais).

Spring tide (coef. 95)

Peak speed ratio model / SHOM: 0.98 (median), range 0.52–1.38. Median phase lag +25 min. Median direction difference nil. 4,104 tides composited.

Neap tide (coef. 45)

Ratio 1.03 (median), range 0.51–1.65. Median phase lag +11 min. Median direction difference 1°. 1,530 tides composited.

One point changes how these figures should be read: SHOM gives a surface current, the model a depth-averaged current. A ratio below 1 — around 0.85 to 0.95 — is therefore expected by construction. The observed medians, 0.98 and 1.03, are not "better" than a perfect match: they indicate that the model runs slightly strong, which errs on the safe side for a stress computation.

The useful discrepancy lies elsewhere. 5 points out of 51 fall outside ±25 % at spring tide. These are coastal headlands — cape accelerations — and banks that the model's roughly 7 km mesh does not resolve. This is consistent with screening use at tile scale, and it is precisely the information to keep in mind for a route that would graze those features: there, the map is not fine enough, and you need to know it before drawing.

Confrontation 2 — the consequence, against 908 observed dune crests

A hydraulic dune is sediment the sea has actually moved. It is the morphological trace of mobility. If the atlas is right, mapped dunes must fall in predicted-mobile areas. So we confront the map with the 908 dune crests from the SHOM "Distribution of dunes on the French continental shelf" product (Open Licence) that fall inside the tile.

The first result is negative: the class-based test carries no information. 31 % of crest-crossed cells are predicted mobile — against a tile base rate of 31 %, a lift of ×1.00. Put differently, predicting "mobile" at random would do exactly as well. On a tide-dominated tile, where the forcing is strong and uniform everywhere, the three-way classification (stable / intermediate / mobile) is too coarse for this test to discriminate. We publish it because a test that fails is information about the domain of validity, not an accident to hide.

The graded test, on the other hand, does carry a signal. At equal sandy substrate — a dune requires sand, and comparing otherwise would compare geologies rather than hydrodynamics — we compare the predicted value at dune cells and elsewhere:

Fraction of time mobile

Median 0.73 at dunes against 0.66 elsewhere. Rank-biserial effect +0.33, p = 7.7 × 10⁻⁹. The signal is clear and significant: dunes do sit where the model predicts more frequent mobility.

Peak stress (90th percentile)

Median 1.36 at dunes against 1.52 elsewhere. Effect −0.08, p = 0.93. No signal.

The honest conclusion is therefore a measured one: it is the frequency of motion that explains where dunes are, not the intensity of stress peaks. That makes physical sense — a dune is built by repetition, not by a single storm — and it tells you which field of the atlas deserves trust when reasoning about morphology.

Finally, the validation is one-sided and must be read as such: an observed dune must fall in a predicted-mobile area, and it does. But the absence of a mapped dune does not prove a stable seabed — survey coverage is partial. The test can confirm; it can never refute.

Confrontation 3 — the weakest assumption: grain size

The threshold of motion depends on the median grain diameter, d50. Yet we do not measure it: we infer it from seabed-nature classes. This is the weakest assumption in the whole chain, and so the one to attack hardest. We confront it with the d50 at the RESOURCECODE hindcast nodes (Krumbein φ scale), which comes from a source independent of the substrate maps. The unit of comparison is the Wentworth grade: one grade = a factor of 2 on d50.

Folk proxy (EMODnet)

913 nodes. Median reference / proxy ratio: 7.85. Agreement within ±1 grade: 29 %. Within ±2 grades: 46 %.

SHOM seabed nature

791 nodes. Median ratio 1.00. Agreement within ±1 grade: 53 %. Within ±2 grades: 75 %.

This is the most useful of the three results, because it is actionable. The EMODnet Folk proxy — the default, free source — is biased by a factor of 7.85 in median on this tile, close to three Wentworth grades. SHOM's 1:50,000 seabed-nature charts show no median bias and agree within ±2 grades in 75 % of cases. Over an area covered by SHOM, there is no reason to settle for the proxy. That is a study-design decision, not an implementation detail, and it comes out of the confrontation — not out of a preference.

What validation changes in the deliverable

A validation is only worth running if it changes what the client receives. Here, it produces three concrete things:

• A robustness map. Replaying the classification with grain size shifted one class finer and then one class coarser, 86 % of nodes keep their mobility class. The remaining 14 % are mapped by name: they are the points where the conclusion depends on the grain-size assumption, and therefore the points where sampling or particle-size analysis is required before design.
• An explicit domain of validity. The model mesh resolves neither headlands nor banks: a route running along them leaves the domain in which the atlas was validated.
• A hierarchy among fields. Mobility frequency is confronted and confirmed; extreme return levels are not confirmed by the dunes, and remain a statistical estimate to be read with its confidence interval.

What we do not claim

These three confrontations validate a screening tool: a first pass at route selection and an aid to scoping, at tile scale. They replace neither a geotechnical survey — cores, CPT, sonar — nor a detailed burial assessment. No atlas built on public data will, and claiming otherwise would be selling a precision that does not exist.

What the atlas does is tell you where the expensive survey should go, and give you the material to defend that choice. The rest — what the atlas does not know — is written in the deliverable, next to what it does.

Are you preparing the lay or burial of a cable, an umbilical or a pipeline? See the Cable corridor atlas offering, or start with the full atlas methodology — from bed shear stress to the A* corridor.

Frequently asked questions