A client asks how accurate the map is. You say 2 cm/px, because that is the number the processing report gave you. The client hears "accurate to 2 centimetres" and puts it in a document that someone will later rely on.
Those are not the same claim. Ground sample distance describes how fine the picture is. Accuracy describes how close a measured position is to the truth. A survey can easily be 2 cm/px and 1.5 metres out, and that combination is the normal result of a drone flight with no ground control and standard GPS.
Key takeaways
GSD is resolution: the ground size of one pixel. It is not an accuracy figure.
Relative accuracy is how good measurements are within the model. Absolute accuracy is how well the model sits on the real world.
With standard drone GPS, expect metre level absolute accuracy however sharp the image looks.
Vertical error is typically two to three times horizontal error. Quote them separately.
Three different numbers, one word
Almost every argument about drone survey accuracy comes from mixing up three separate quantities. It helps to name them before using any of them in a sentence to a client.
Number | What it describes | Typical driver |
|---|---|---|
Ground sample distance | The ground size of one output pixel | Flight height, sensor, focal length |
Relative accuracy | Error between two points inside the same model | Overlap, image sharpness, reconstruction geometry |
Absolute accuracy | Error between a point in the model and its true coordinate | Georeferencing method: GPS, RTK, PPK, ground control |
A stockpile volume depends almost entirely on relative accuracy: the pile and its base are both inside the same model, so a shared offset cancels out. Setting out a foundation from the same map depends on absolute accuracy, because the coordinates have to agree with the real world and with everyone else's drawings. Same file, completely different requirement.
What GSD actually guarantees
GSD guarantees one thing: what you can see. At 2 cm/px, a 10 cm wide painted line is five pixels across and unambiguous. A 1 cm crack is invisible, whatever the accuracy of the georeferencing.
It also sets a floor on what you can measure. You cannot digitise an edge more precisely than roughly one pixel by eye, and in practice a careful operator lands somewhere between one and two pixels. So a 2 cm/px map supports edge picking at the two to four centimetre level, before any georeferencing error is added on top.
That floor is a good sanity check on a deliverable. If somebody reports a length to the millimetre from a 3 cm/px orthomosaic, the extra digits are decoration.
Where absolute accuracy comes from
The model itself does not know where it is. Photogrammetry reconstructs a self consistent shape from the photos, and that shape is then placed in the world using whatever position information is available. The quality of that placement is your absolute accuracy, and it depends almost entirely on the method.
Georeferencing | Typical horizontal | Typical vertical | Notes |
|---|---|---|---|
Standard drone GNSS from EXIF | 1 to 5 m | 2 to 10 m | The default for most consumer drones |
RTK or PPK on the aircraft | 2 to 5 cm | 3 to 8 cm | Needs a base station or a correction service |
Surveyed ground control points | 2 to 5 cm | 3 to 6 cm | Independent of the drone, needs field work |
RTK plus a few checkpoints | 2 to 4 cm | 3 to 5 cm | Best practice: control and verification |
Two things about this table are worth saying out loud. First, the jump from the first row to the rest is two orders of magnitude, and no amount of processing quality closes it. Second, the numbers in the lower rows are achievable, not automatic: badly placed ground control, a mistyped coordinate or a base station on an assumed position all produce confident, precise, wrong results.
Ground control points, briefly
A ground control point is a marked target on the site whose coordinates you measure independently, usually with a GNSS rover or a total station. The reconstruction is then fitted to those coordinates. Five to ten well distributed points across a site is a common starting figure: spread them across the full extent, include the corners, and add a couple in the middle rather than clustering them along one edge.
Keep two or three of them out of the fit as checkpoints. Points used to fit the model cannot also verify it. The residual on a checkpoint that the solution never saw is the only honest accuracy statement you can make.
See what your own photos say about the flight
Every processed project keeps a record of how it was made: photos used, estimated GSD, the GPS accuracy declared by the camera, whether RTK was fixed, and the output resolution requested. The free plan takes up to 150 photos per project.
Process a test flightVertical accuracy is the one that bites
Photogrammetry solves height from the intersection of rays that come from above, at fairly shallow angles to each other. That geometry is weaker vertically than horizontally, which is why the vertical error is routinely two to three times the horizontal error on the same survey.
It gets worse in three specific situations, all of which are common:
At the edges of the flight. Points near the boundary appear in fewer photos and from fewer angles. Elevation there is noisy and can be visibly warped. This is why you fly a buffer beyond the area of interest.
On flat, uniform surfaces. Fresh asphalt, calm water, uniform sand and large uninterrupted roofs give the matcher nothing to lock onto. The surface can bulge or dish without any obvious sign in the image.
Under vegetation. The bare earth model there is interpolated between whatever ground points exist at the edges of the canopy, so it is an estimate.
There is also a systematic vertical trap that has nothing to do with photogrammetry: the difference between ellipsoidal and orthometric height. Drone GNSS reports height above the ellipsoid. Site drawings almost always use height above a local vertical datum. The offset between them can be tens of metres, and it is constant, so it looks nothing like noise. It looks like a perfectly good survey that is thirty eight metres out.
How to state accuracy without overpromising
The safest delivery note is three sentences long and separates what you know from what you assume.
State the resolution: "the orthomosaic has a ground sample distance of 2.4 cm/px".
State the georeferencing method: "the survey is georeferenced from the aircraft's onboard GNSS, without ground control".
State the expected absolute accuracy as a range, and say how it was established: "expected absolute accuracy is in the metre range horizontally and worse vertically; no independent checkpoints were measured".
If you did measure checkpoints, replace the third sentence with the actual residuals. A single line saying "checked against three surveyed points, maximum residual 3.1 cm horizontal and 4.6 cm vertical" is worth more to a client than any generic claim, and it protects you if a dispute follows.
A checklist before you quote a number
Is the client asking about relative measurement inside the model, or absolute position in their coordinate system? Ask, do not guess.
Which EPSG code do their existing drawings use? Deliver in that, not in whatever the software defaulted to.
Which vertical datum? If nobody knows, deliver ellipsoidal heights and label them clearly.
Were there checkpoints? If not, say so.
Was the flight flown high over a flat, featureless surface? If so, expect the vertical to be the weak part.
Are you comparing two surveys over time? Then relative consistency between flights matters more than either flight's absolute accuracy, so keep the flight parameters identical.
Frequently asked questions
Is 2 cm/px the same as 2 cm accuracy?
No. 2 cm/px means each pixel covers 2 cm of ground, which is a statement about detail. Accuracy is how close a measured coordinate is to the truth, and with standard drone GPS it is usually in the metre range regardless of how fine the pixels are. The two numbers are independent and should be quoted separately.
Do I need ground control points if my drone has RTK?
For georeferencing, usually not. For verification, yes. RTK gives you centimetre level positions but no independent evidence that the result is correct, and RTK setups fail quietly: a wrong antenna height, a base on an assumed position, or a dropped fix mid flight. Two or three surveyed checkpoints turn an assumption into a measurement.
Why is my elevation off by tens of metres when the map looks perfect?
Almost certainly a vertical datum mismatch. GNSS reports ellipsoidal height, while site drawings usually use an orthometric height above a local datum. The difference is a large, constant offset. It has to be corrected with a geoid model or a local shift, not by reprocessing.
Does higher processing quality improve accuracy?
It improves detail and noise, not georeferencing. A denser reconstruction resolves finer features and gives a cleaner surface, which helps relative measurement. It cannot move the model onto the correct absolute coordinates, because that information is simply not present in the photos.
Accuracy is not one number and it is not a property of the software. It is the combination of how fine you flew, how well the reconstruction held together, and how the model was tied to the real world. Say which of the three you are quoting, quote horizontal and vertical separately, and back the claim with checkpoints whenever the client's decision depends on it. If you want to see how the plans differ on processing quality and how long results stay available, that is on the pricing section.