Your RTK Drone Flew Perfectly. Without Checkpoints, You Still Can’t Prove the Survey Is Right.

Your RTK Drone Flew Perfectly. Without Checkpoints, You Still Can’t Prove the Survey Is Right.

A client looks over your specification and puts forward a reasonable question. Since your drone has RTK built in, why are you still quoting ground control points? Doesn’t the drone already know its position?

It often does so; there have been a number of studies in which RTK drones were flown without any GCPs at all and yet still managed to land within a few centimetres of the independent checkpoints. Therefore, I would not say today that you always need control points.

But the unpleasant thing is this: if a survey that relies only on RTK is incorrect, it seldom appears to be wrong. The orthomosaic remains sharp, and the processing report is clean. A height offset or a gradual tilt over the site stays undetected until someone compares the surface to a known point; that person is usually the client’s engineer, and the discovery typically occurs after the invoice has been paid.

The actual issue isn’t RTK or GCPs; it’s whether you can prove the result in either case. Let me clarify the way I see this.

What RTK actually fixes (and what it doesn’t)

GCPs have two functions: they attach the project to real-world coordinates, and they keep the reconstructed geometry in shape so that the model does not drift or bend. That is how RTKdata’s engineering analysis of GCP-free mapping puts the two functions apart, and it is the essential point of the entire subject. RTK carries out the first function but does not perform the second.

In the case of a fixed solution, the receiver has knowledge of the position of the antenna to within a few centimetres, and the processing software uses this information to position each image. That’s fine. However, with RTK, the antenna is positioned, not the camera. The antenna is located on the airframe while the camera is lower down, and this lever arm is often between 15 and 20 cm, although it can be greater. If this offset is not taken into account, all the images will have the same systematic error even though the RTK has carried out its function perfectly.

Timing is also important; since a drone travelling at 8 m/s passes 16 cm in 20 milliseconds, a small delay between the GNSS timestamp and the shutter can result in a loss of several centimetres. On integrated platforms, the manufacturer usually looks after both of these aspects. In the case of retrofitted or third-party payload installations, I would pay more attention.

A few words regarding RTK and PPK, since the two are often confused. RTK applies the corrections in real time and requires a stable connection with a base station or network, whereas PPK records the raw observations and then applies the corrections afterwards, which means it does not rely on the link remaining intact during flight. Neither method deals with the issues that occur later on.

What the studies show when you remove the GCPs

Begin with the positive aspect. In a 2020 study published in the journal Drones, the researchers flew a DJI Phantom 4 RTK over a 2 km by 0.5 km section of a braided river in Scotland, the terrain having a relief of about one metre. When no ground control points were used, the vertical RMSE was 6.6 cm; with five ground control points, it was 7.2 cm, so the use of the control points did not result in any real improvement in this case. They also detected no doming regardless of whether or not the control points were used. A 2024 paper in the ISPRS Annals used the same drone model with 71 spatially distributed checkpoints and no ground control points and recorded a vertical RMSE of 5.58 cm.

Take that and compare it with the situation before the introduction of RTK. A previous Korean study comparing drones found that a non-RTK drone without GCPs was about 113 cm off horizontally and 205 cm off vertically, whereas a drone using RTK without GCPs was only about 13 cm and 16 cm off, respectively. That degree of improvement is what RTK provides, and it’s far from negligible.

On the other side, drone testing carried out by DroneDeploy and published by Skydio compared a number of positioning methods. Using only RTK and PPK did produce usable results, although with greater variability—particularly in the vertical direction—and the inclusion of GCPs greatly reduced the total error. Since this is vendor testing, it should be taken with caution, but it still supports the findings from the independent studies.

Here are two points to note. Firstly, under favourable conditions and mainly with one drone family, the good results are about 5 to 7 cm vertically in terms of RMSE. This makes the results suitable for mapping and monitoring. It is a separate question whether this level of accuracy is adequate for a design surface that requires a tolerance of 3 cm.

Secondly, all of these figures exist only because some independent checkpoints were surveyed. Although the 2020 team was satisfied with its no-GCP result, it still advised that checkpoints be spread out. That is indeed the main point of the post.

Where the gamble hides

Heights that are the wrong kind of height

RTK provides ellipsoidal height, which is measured from a mathematical reference surface. Most customers require orthometric height, this being linked to a geoid model or a national vertical datum. The two can differ by as much as tens of metres, and the difference varies from location to location. In certain areas of Europe, the separation exceeds 45 m. If the wrong geoid model is used or no model at all is applied, your heights will be incorrect even though the horizontal position appears to be accurate.

Since the degree of separation differs from one point to another on a site, the error may appear as a tilt rather than a clear offset, and it is hard to detect without using independent checkpoints. We discussed the horizontal aspect of this issue in our post on CRS errors in drone deliverables.

A base or datum you haven’t verified

The network corrections are associated with a specific reference frame and epoch, whereas your final product might require a national datum, a project datum, or one that is older. In areas that are tectonically active, the frames can drift apart by as much as a few centimetres each year. A project reference frame that is five years old could therefore be about 10 cm away from one that is given in the current frame. If you set up your own base station, then the coordinates of that base station become the starting point for all the other measurements, and an incorrect position for the base station will affect the entire survey.

The most unpleasant one is that one, since all the values within the survey remain consistent. The distances and areas are still measured accurately. The only thing that is wrong is the absolute position, and no one picks it up until your data is compared with existing control. If you want more details about the epoch side, refer to the explanation regarding why repeated drone surveys don’t align.

A fix that quietly wasn’t one

An RTK solution may switch from fixed to float mode due to a long baseline, poor mobile coverage, or the presence of multipath close to buildings. As a result, lower-quality position data can be included in the image block and can affect the adjustment. When you record the raw observations on board, PPK provides an option for recovering that flight rather than having to refly it. In either case, you should check the solution status for each image before accepting the positions.

Doming on a big nadir-only block

If the camera is pointed straight down and a large block is being flown, the errors in camera calibration and block geometry can balance one another, and the surface will bend slightly. Although accurate image positions lessen this risk, they do not eliminate it.

It doesn’t have to involve a field full of GCPs; oblique imagery, cross strips, or one or two separate checkpoints could be used to verify the result. And it’s worth remembering that the 2020 study found no doming at all, so it by no means has to occur. It should be checked rather than assumed.

Four questions to ask before you skip the GCPs

Question RTK-only is a reasonable plan when… Add GCPs (or more control) when…
How great a tolerance does the deliverable require and who is responsible for auditing it? General mapping, progress monitoring and documentation, tolerances of several centimetres Design surfaces, contract-sensitive or dispute-prone deliverables, tolerances close to those of the RTK-only results above
To what extent does height play a role? Primarily planimetric work Grading, drainage, and volumes—any work that is based on Z
How weak is the flight geometry? Moderate area, the presence of cross strips or oblique images, and varied terrain Large nadir-only block, which is either repetitive or of low texture
Shall you place your trust in the base and datum chain? The base is based on a known point, the geoid model has been confirmed, and the frame agrees with the deliverable Unknown base, mixed datums, tying to the existing project control

This isn’t a set of rules, and the contract or specifications always take priority. However, if two or more of your answers end up in the right-hand column, I would recommend laying out control. If most of your answers are in the left-hand column, then an RTK-only solution with checkpoints is a reasonable approach, and I would include that approach in the report.

Control less, check more

The reason why an RTK-only method can be regarded as valid is that control points are employed to fit the model while checkpoints are kept back and used to test it. After a point has been used to modify the model, it can no longer truthfully indicate how good the model is. Therefore, an RTK-only workflow can be rigorous provided that the field time saved by not having to visit the targets is allocated to checkpoints.

What is the number? For formal accuracy assessment, the second edition of the ASPRS positional accuracy standards increased the minimum number of checkpoints from 20 to 30. Although most commercial jobs are not subjected to such a formal assessment, the principle can still be applied. When carrying out measurements, distribute them throughout the corners, over the lowest ground and the highest ground, and not merely in areas near the access road. For a more in-depth look at why QA/QC is the real bottleneck in drone mapping, refer to our post on incorporating this into a workflow.

We carried out a survey at a dam site in a mountain river valley using a fixed-wing drone which flew at a height of 99 m, capturing 403 images of an area amounting to 0.317 km², the images including terraced farmland, cliff faces and an active watercourse. The survey was linked to 12 ground control points that were observed using RTK/GNSS, and the total RMSE reported was 5.5 cm. In a location that is either difficult or unsafe to access on foot, I prefer to have control in the model rather than rely on the positions of the images alone. Indeed, the figure only has any meaning since it is based on points that have been surveyed.

Is it possible to do without GCPs when using an RTK drone?

In some cases, yes. The aforementioned studies indicate that RTK-only can achieve a vertical RMSE of about 5 to 7 cm on independent checkpoints under favourable conditions. It is possible to justify skipping control. However, you cannot justify skipping the checks.

Is PPK superior to RTK?

They are solving different problems; PPK does not require a live link and provides a recovery option, whereas RTK gives you positions while in flight. Neither of them corrects the geoid, the datum, or poor flight geometry, so use both if your hardware permits it.

How many checkpoints do I need?

Formal requirements specify a minimum of 30; for an actual job, a smaller number of well-distributed points gives you a lot more information than having none at all. The only thing you should not do is use them to fit the model.

What position does this leave you in?

To cut to the chase, RTK transferred the work rather than eliminating it. The time you save by not setting targets must be accounted for: it has to be allocated to the lever arm, the geoid model, the base, and a number of checkpoints. Once you can show your client those four items, then using RTK alone is a fully defensible approach to surveying. But if you can’t do that, you’re asking them to accept your word for it.

When considering whether to choose RTK-only or go with full control as part of a live project, our drone imaging and processing team would be more than happy to go over the trade-offs.

Back to blog

Leave a comment