Key takeaways
- An accuracy number means nothing until it says absolute or relative. Relative accuracy describes measurements taken within the model; absolute accuracy describes how well the model agrees with the real coordinate system your project works in.
- Horizontal and vertical are separate numbers and vertical is almost always the weaker of the two. A single figure quoted for both is a sign nobody has checked.
- RMSE is an average of errors across checked points, not a worst case. Half the sample sits above it by definition.
- A 95 percent confidence statement is roughly 1.96 times the RMSE for vertical. A quote giving RMSE and a quote giving 95 percent confidence can describe the same survey with numbers that differ by about double.
- Ground sample distance is pixel size, not accuracy. A 1 cm pixel on a job with no ground control can still sit half a metre from where it belongs.
- The only evidence of absolute accuracy is a comparison against surveyed points the processing never saw. Residuals on the control points describe the fit, not the truth.
- Put the test in the purchase order: name the datum, name the tolerance, say absolute, and require a check point table in the deliverable.
Open three drone survey quotes for the same site and you will usually find three accuracy numbers, none of them defined. One says 2 cm. One says sub-centimetre. One says survey grade. They are not comparable, and at least one of them is almost certainly describing something other than what the buyer thinks it describes.
This is not usually dishonesty. Accuracy in photogrammetry is several different measurements that share one word, and a quote has room for one number. The problem is that the cheapest of those measurements to achieve is also the one that sounds best, so the number that gets quoted is often the least useful one.
Here is what the figure has to say before it means anything, and the one test that settles it.
Absolute accuracy and relative accuracy are different claims
This is the distinction that matters most and the one most often left out.
Relative accuracy describes how well measurements taken inside the model agree with reality. The distance between two points on the ortho, the height of a pile from its toe to its crest, the width of a pad. The model is internally consistent, and a measurement taken wholly within it is close to correct.
Absolute accuracy describes how well the model agrees with the coordinate system and vertical datum your project actually works in. It answers a different question: if I hand this surface to the engineer, will the design elevations line up, and will it tie into the existing survey.
A flight with no external reference at all can produce excellent relative accuracy and be metres out in absolute terms. The model is the right shape, sitting in the wrong place. Nothing in the imagery looks wrong, because every feature is in the right position relative to every other feature.
So the first question on a quote is not how many centimetres. It is centimetres of what. If the answer is relative accuracy and the deliverable is going to a designer or into a project survey, the number is not answering the question you need answered.
Horizontal and vertical are two numbers, and vertical is the hard one
Photogrammetry derives height from the geometry between overlapping images, and that geometry is weaker in the vertical axis than the horizontal. On most jobs the vertical figure is one and a half to two times the horizontal one, and a surface used for volumes or grading lives or dies on the vertical.
A quote giving one number for both has either not separated them or is quoting the horizontal because it reads better. Ask for both.
Vertical is also where the systematic failure hides. GNSS measures height against an ellipsoid, and a project works in orthometric heights, so a geoid model converts between them. An error in that conversion, or the wrong geoid, or a mismatch between the datum the contractor used and the one the project is on, shifts the entire surface up or down by a constant amount. That offset does not appear anywhere in the imagery and does not degrade the model's internal quality at all. It changes every elevation and every volume on the job. The coordinate system named on the deliverable is worth confirming before the flight, not after.
What RMSE is actually telling you
Most honest accuracy figures are root mean square error, and RMSE is widely misread as a maximum.
It is not. RMSE takes the differences between the model and a set of checked points, squares them, averages the squares and takes the root. It is a measure of typical error across the sample. By construction, individual points sit above it. A survey reporting a vertical RMSE of 3 cm will contain points off by 6 cm, and that is normal rather than a defect.
Two further things the number does not carry on its own:
- The sample size and where the samples were. An RMSE computed from four points near the parking area describes the parking area. Error grows toward the edges of the photo block and in the parts of the site with the least control, so where the checked points sat is part of the result.
- The confidence level. Positional accuracy standards for geospatial data are commonly expressed at the 95 percent confidence level, which for vertical accuracy is approximately 1.96 times the vertical RMSE. That matters when comparing quotes, because a contractor quoting raw RMSE and a contractor quoting accuracy at 95 percent confidence can be describing the same survey with figures that differ by roughly a factor of two. The one with the bigger number may be the more rigorous operator.
So "2 cm" needs a third qualifier alongside absolute and vertical: 2 cm as what. RMSE, or 95 percent confidence.
Ground sample distance is not accuracy
Quotes often lead with a resolution figure, and it is easy to read as an accuracy claim. It is not one.
Ground sample distance is the real-world size of one pixel. A 1 cm GSD means one pixel covers one centimetre of ground, which governs the smallest feature you can see and how crisply an edge resolves. It says nothing about whether that pixel is in the right place.
Resolution and positional accuracy come from different parts of the job. GSD comes from the camera and the flight altitude. Absolute accuracy comes from RTK or PPK positioning, from ground control, from the datum handling and from the processing. A beautifully sharp 8 mm orthomosaic flown with no external reference is a high-resolution picture in an unknown location.
A useful rule of thumb for a well-controlled job is that absolute accuracy lands in the range of a small multiple of the GSD, not at the GSD itself. If a quote implies accuracy equal to pixel size, the number is optimistic.
How to actually test the claim
Everything above is how to read the number. This is how to find out whether it is true, and it is simpler than it sounds.
Accuracy is evidenced by check points: surveyed targets that are deliberately withheld from the processing solution, so the software never sees them. After the model is finished, its value at those locations is compared against the surveyed values, and the differences are the result. That comparison is the only independent evidence of absolute accuracy that exists on a drone job.
What is not evidence is the residual table in the processing report. Those residuals describe how closely the model fits the control points it was explicitly told to match. A model can report tight residuals on its control and be wrong everywhere between them. Reporting control residuals as an accuracy figure is the single most common way an accuracy claim ends up overstated without anyone lying, and how the control and check points are split is the part of the field plan that decides whether you get real evidence.
So the test is one question: can you show me the check point comparison, from points that were held out of the solution?
Three possible answers. A table of held-out points with their differences, which is the right answer. An offer to add check points and report them, which is a fair answer on a job that was not scoped that way. Or a restatement of the control residuals as though they settled it, which tells you the accuracy figure on the quote has never been independently tested.
Put the test in the purchase order
An accuracy requirement only binds if it is specific. Four lines are enough:
- Name the datum and coordinate system. Horizontal and vertical, including the geoid model. Without this the tolerance has nothing to be measured against.
- State absolute, and give horizontal and vertical separately. For example: absolute horizontal within X, absolute vertical within Y, in the named datum.
- Say which statistic. RMSE or 95 percent confidence. Pick one so every bidder prices the same thing.
- Require the check point table in the deliverable. A named number of independently surveyed points, withheld from processing, reported with their differences.
Those four lines do more for comparability than any amount of negotiating over the headline figure, because they force every quote onto the same definition. They also change what you are buying: the fourth line is what turns an accuracy claim into an accuracy record, which is what matters if the deliverable is ever relied on by someone who has to sign for it.
Not every job needs the tightest number
Specifying more accuracy than the deliverable needs is its own waste, because tightening the tolerance adds ground control, field time and processing to the price.
- A visual progress record or a marketing orthomosaic. Relative accuracy is sufficient. RTK or PPK alone, no control campaign.
- A stockpile volume for internal reporting. Good relative accuracy and a verified vertical. The volume is a difference between two surfaces, so a uniform offset mostly cancels, but the base surface has to be right.
- A volume that settles a payment or an inventory audit. Absolute, verified, with held-out check points and the comparison in the report.
- A surface going into design, or tying into an existing project survey. Absolute in the project datum, vertical specified separately and tightly, check points mandatory.
The job is to match the specification to the decision the data supports, then make the number on the quote say what it means. A contractor who answers the absolute-or-relative question without hesitating, gives you two numbers rather than one, names the statistic and offers the check point table is telling you more about the survey than the figure itself ever could.