Auditing a wave model: ERA5 and MFWAM against 11 altimetry missions and CANDHIS buoys

Why audit a wave model before relying on it

ERA5 (ECMWF) and MFWAM (Copernicus) produce a significant wave height at every point, every hour, going back decades. That is what makes them indispensable for sizing a cable-lay operation or deciding a weather window. It is also what makes them dangerous to take at face value: a 0.5° reanalysis grid does not necessarily "see" the coastline, and a systematic bias of a few tens of centimetres shifts the exceedance probability of an operational threshold.

We built a four-phase audit across three corridors — the English Channel (cable corridor), the Gulf of Lion, and the Raz de Sein — covering 2010 to 2026, with a structural gap between 2021 and 2022: the multi-year archive for some Copernicus missions ends in June 2019, and the near-real-time (NRT) archive starts later. This gap cannot be filled with current Copernicus products — we flag it rather than interpolate over it silently.

Phase 1 — Jason-3 against ERA5, over the Channel

First judge: the Jason-3 altimeter, compared point by point against ERA5 over the Channel corridor. The global bias is −0.066 m, for an RMSE of 0.332 m and a scatter index (SI) of 17.6%. This global figure hides a heterogeneity worth naming:

Open water (32.4%, n = 7,848)

Bias −0.057 m, RMSE 0.287 m, SI 15.6%.

Coastal zone (67.6%, n = 16,432)

Bias −0.072 m, RMSE 0.359 m, SI 19.2%.

Two-thirds of the matchups over the Channel are coastal — precisely where nadir altimetry itself loses quality near land. Treating this population separately, rather than diluting it into a single global average, is what makes the number usable.

The seasonal RMSE is tight: 0.345 m in winter (DJF), 0.343 m in spring (MAM), 0.283 m in summer (JJA), 0.343 m in autumn (SON) — summer is the only season clearly apart, consistent with calmer seas. The most operationally useful signal comes from the QQ-plot: it shows tail compression. ERA5 underestimates extremes above roughly 2 m Hs — exactly the range that triggers standard operational thresholds.

Phase 2 — eleven missions, to check the bias is stable

A bias measured on a single mission says nothing about its stability: is it specific to the sensor, or structural in the model? We extended the comparison to eleven altimetry missions (Jason-3, Sentinel-3A, Sentinel-3B, Sentinel-6A, SARAL, CryoSat-2, and earlier missions) over the Channel. Inter-mission bias differences range from 0.04 to 0.13 m — consistently below the altimetric uncertainty floor itself, 0.2 to 0.3 m.

Operational conclusion: there is no differential bias between missions, so no per-sensor recalibration is needed. The MY-to-NRT archive transition — from Jason-3 to Sentinel-6A, in April 2022 — does not break bias stability despite the change of platform. One exception to note: CryoSat-2 is only usable over 2010–2013 (n = 101) before its switch to SAR mode, which changes the nature of the measurement.

Phase 3 — three corridors, one starting assumption contradicted

Extending the audit to the Gulf of Lion and the Raz de Sein produces two notable findings. First, ERA5's 0.5° resolution is too coarse for the Gulf of Lion — a corridor where bathymetry and coastline vary at a scale the grid does not capture. Second, and this contradicts our starting assumption: the Raz de Sein is not the worst-performing corridor. We expected a narrow, exposed passage to penalise the reanalysis the most; the comparisons do not show that.

Over that same Raz de Sein, MFWAM (Copernicus, 0.083° resolution, roughly 9 km) offers a finer alternative, but with a short time window: November 2022 to June 2026, with only four near-real-time missions available over that period. The finer resolution therefore comes at the cost of a history reduced to 3.6 years.

Phase 4 — CANDHIS buoys as a third judge

Satellite altimetry measures along a ground track, at a given instant. Cerema's CANDHIS network — around a hundred coastal buoys spread along the French coastline — measures at a fixed point, continuously. It is a judge of a different nature, which is what makes it a complementary test rather than a redundant one.

Over the Raz de Sein, comparing ERA5 against altimetry near a dozen CANDHIS stations yields 21,603 matchups, for a mean Hs of 2.20 m — consistent with the phase 3 findings.

The most important result of this phase is an instructive failure, not a number: over the Gulf of Lion, all ten CANDHIS buoys fall into cells that ERA5 masks as land, due to its 0.5° grid. Validation is simply impossible there with ERA5 — this is not a question of model quality, it is a resolution problem that blinds the model exactly where the instruments sit. Switching to MFWAM puts four of the ten buoys — notably Le Planier and Espiguette — back at sea and makes them comparable: 109,283 matchups over January 2022 to June 2026, for a bias of −0.043 m, an RMSE of 0.183 m, and an SI of 23.3%.

What the audit changes, and what it does not claim

Three reading caveats apply, and we publish them alongside the numbers rather than in a footnote:

• Bias differences of 0.04 to 0.13 m between missions, as in phase 2, are below the uncertainty floor of altimetry itself (0.2–0.3 m): they should not be read as real model error.
• Nadir altimetry degrades near coasts: 67.6% of the Channel matchups are coastal and must be read separately from open water, never blended into a single average.
• The MFWAM archive spans only 3.6 years, and a CANDHIS buoy measures at a point while a model averages over a grid cell: the two judges complement each other, neither replaces the other.

The audit does not say "the model is good" or "the model is bad." It says where it is usable — the mobility frequency signal from phase 1 holds, the inter-mission stability from phase 2 holds — and where it is structurally blind, such as a 0.5° grid that places ten buoys on dry land. It is this reliability map, not a single average, that belongs in a weather window or a design threshold.

Frequently asked questions