A geoid model is a gridded surface of geoid heights, N, covering a region. You interpolate it at your latitude and longitude, and it hands back the separation between the ellipsoid and the geoid at that point so you can convert a GNSS height into an elevation. That is the entire function.
What makes the topic worth a page is that there are two fundamentally different kinds of geoid model, they are built for different purposes, and using the wrong one produces a plausible-looking elevation that is wrong by a decimetre or more.
Gravimetric versus hybrid
- Gravimetric geoid model
- A model of the true geoid computed from gravity observations — terrestrial, airborne and satellite — together with terrain data. It is a physical model of the Earth gravity field and knows nothing about any particular national datum.
- Hybrid geoid model
- A gravimetric model warped to fit the existing national vertical datum, so that a GNSS ellipsoid height minus the modelled geoid height reproduces the published elevation on the existing benchmark network. It is a conversion tool, not a physical model.
GEOID18 is a hybrid model. It was fitted to GPS-on-benchmark observations so that ellipsoid heights on NAD 83 (2011) convert to published NAVD 88 elevations. That fitting is what makes it useful today and also what makes it a bridge to a datum with a known half-meter bias and continental tilt — the model faithfully reproduces those defects, because reproducing them is its job.
The succession of models
| Model | Type | Fitted to |
|---|---|---|
| GEOID09 | Hybrid | NAD 83 (NSRS2007) and NAVD 88 |
| GEOID12A / GEOID12B | Hybrid | NAD 83 (2011) and NAVD 88 |
| GEOID18 | Hybrid | NAD 83 (2011) and NAVD 88, with additional GRAV-D airborne gravity |
| Experimental xGEOID series | Gravimetric | No national datum — the true geoid, used to develop the modernized datum |
| Model for the modernized datum | Gravimetric | The geopotential datum itself, not a benchmark network |
The trajectory is clear enough. Each hybrid generation incorporated more gravity data and fitted a better GPS-on-benchmark dataset, while a parallel experimental line developed the purely gravimetric model that will define the modernized vertical datum. When that arrives, hybrid models become unnecessary, because the geoid model is no longer a bridge to a benchmark network — it is the datum.
What accuracy to expect
A current hybrid model in the conterminous United States is generally good to a couple of centimetres in flat, well-surveyed terrain, degrading in the mountains and near the edges of the data. Alaska is weaker than CONUS because both the gravity coverage and the benchmark network are sparser.
Two distinct error types are worth separating. There is model error — the difference between the model and the true relationship — which is generally small and smooth. And there is local datum error: the benchmark you are comparing against may itself be mispositioned in NAVD 88, or may have physically moved. The second is often larger than the first, and it looks identical in your residuals.
This is why the standard procedure is to occupy several published benchmarks distributed around the project, compare modelled against published elevations, and inspect the residuals as a set. Consistent residuals across all of them indicate a local datum offset you should absorb. One outlier among several consistent marks indicates a disturbed benchmark you should reject.
Using a model correctly
- Confirm which model your data collector or processing software is applying. Defaults are frequently out of date.
- Confirm the ellipsoid heights going in are on the realization the model expects.
- Interpolate rather than snapping to the nearest grid node. Geoid height changes by centimetres per kilometer, so nearest-node interpolation over a coarse grid introduces steps.
- Hold at least two, preferably three or four, published benchmarks bracketing the project.
- Record the model name on the deliverable alongside the vertical datum.
- Do not average two different geoid models in an attempt to split the difference.
Why the gradient matters more than the value
Most projects are not sensitive to an absolute geoid height error, because a constant offset is absorbed the moment you hold a single benchmark. What projects are sensitive to is the gradient — the way N changes across the site. An error in the gradient tilts your entire surface.
For a small site, a single value of N is often adequate precisely because the gradient over a few hundred meters is negligible. For a long linear project — a highway corridor, a levee, a pipeline, a runway approach — it is not. Geoid height can change by a decimetre over ten kilometers in ordinary terrain, and faster in the mountains, and that shows up directly as a false grade. Corridor projects should hold benchmarks at both ends and in the middle, not at one end.
The same reasoning explains why the modernized system invests so heavily in gravity. A geoid model that is right in an absolute sense and right in its gradients removes the need to level long distances at all, which is the point of the whole exercise.