Drone survey accuracy: what GCPs, RTK and PPK really give you

Accuracy claims for drone surveys are usually a single number that hides three different questions: how fine is the map, how close is each point to where it really is, and how do you know. This guide separates them, gives the numbers you can expect with each positioning method, and ends with the one test that makes a claim into a fact.

Three different things called accuracy

• Resolution is the ground sample distance, the size of one pixel on the ground. A 20 MP camera at 100 m gives about 2.7 cm; at 60 m, about 1.6 cm. Resolution bounds accuracy but is not accuracy.
• Relative accuracy is how well the model's shape matches the ground: the distance between two points on the map against the distance on the ground. Photogrammetry is good at this. A flight with no positioning at all can have the right shape to within a few centimetres.
• Absolute accuracy is how close a point on the map is to its true coordinate in the world. This is what positioning methods buy, and what you need to compare two flights, to lay the map on a plan, or to set out from it.

What each positioning method gives you

Drone GPS only

The positions written into the photos come from the drone's ordinary receiver, good to a metre or so horizontally and worse vertically. The photogrammetry still builds a correct shape from the overlap; it just sits that shape where the drone thought it was, and scales it from those positions, so the model can be a percent or two large or small. Good enough for a volume of one pile on its own, for a look at the site, for an orthomosaic to point at. Not good enough for anything that has to agree with another survey.

Ground control points

Targets laid on the ground and measured with a survey-grade receiver before the flight. The software finds each target in the photos and pulls the model onto the measured coordinates, correcting position, scale and tilt at once. Five is the practical minimum; more, well spread, is better; a site with steep relief wants some on the high ground and some on the low. GCPs cost a visit to every target with a rover, which is the main reason RTK and PPK exist.

RTK

Real-time kinematic: the drone receives corrections from a base station during the flight and records each photo's position to a few centimetres as it is taken. The catch is the word real-time. If the radio or mobile link to the base drops for a leg of the flight, those photos revert to ordinary GPS, and a model that mixes fixed and unfixed positions can be worse than one with neither. Check the fix status in the flight log before you trust it.

PPK

Post-processed kinematic: the drone records its raw satellite observations during the flight, and the positions are solved afterwards against a base's observations of the same satellites. Nothing has to be received in the air, so there is no link to drop; the base can be a public reference station that was never on the site; and the solution can run both forward and backward through the flight, which makes it the more reliable of the two. The cost is a step after landing. Clarity does that step itself: drop the drone's satellite files in with the photos, choose the nearest NOAA reference station (in the United States) or your own base file, and the corrected positions are solved before processing starts, with the station, the distance and the number of fixed photos printed on the survey. (How processing with PPK works.)

The vertical is always worse

Every method is roughly twice as good horizontally as vertically. Satellites are overhead, so geometry constrains height less; photogrammetry's depth comes from parallax between frames, which is weaker than the position within a frame. Published standards reflect this: a survey quoted at 3 cm horizontal is usually 5 to 8 cm vertical. For earthwork, the vertical is the one that matters, so that is the number to ask for.

How to know: checkpoints

An accuracy figure without checkpoints is a specification, not a measurement. The test is simple. Lay a few targets the model is not told about, measure them with a rover, and when the map is finished, read each target's position off it. The differences are the error; the root mean square of them is the figure standards bodies use. Three well-placed checkpoints tell you more than any datasheet. Keep the comparison with the survey, because when two flights disagree, it is the first thing anyone will ask for.

What accuracy means for the numbers you use

• A single volume needs scale, not position. Any method, even GPS only, gives a volume within a few percent; the base under the pile decides the rest. (Measuring a stockpile.)
• Cut and fill against a design needs the survey on the design's datum, to a few centimetres vertically. Over a hectare, 3 cm is 300 m³. Control, RTK or PPK, and a check. (Cut and fill methods.)
• Two flights compared need the same datum twice. The commonest failure is not noise but a whole-site offset of a few centimetres to a metre between flights that were each fine on their own. Clarity reports that offset when it sees one and lets you level the pair on unchanged ground. (Comparing two surveys.)