"Near Survey Grade": What a Non-RTK Drone Actually Delivers, With Numbers

Search for drone mapping in Arizona and you will find a lot of pages claiming survey-grade accuracy. Very few of them name the aircraft. Almost none mention ground control. Not one that we found states what the accuracy becomes when you skip it.

We would rather publish the whole picture, including the unflattering half, because anyone who actually needs accurate data will ask these questions eventually and it is better to answer them on a web page than in a dispute.

The aircraft, and why it matters

We fly a DJI Mavic 4 Pro. It is an excellent camera platform — a 100MP 4/3 Hasselblad main sensor and 6K video.

It is not an RTK aircraft. There is no RTK module for this airframe, it is not an enterprise variant, and it does not log the raw GNSS observables that post-processed kinematic workflows need. Its positioning is single-frequency consumer GNSS.

That single fact determines everything below, so it goes at the top rather than in a footnote.

The numbers

Relative accuracy — measurement between two points inside the model, which is what most people actually need — is better than absolute, because it is unaffected by datum offsets. Point-to-point is typically 2–6 cm, linear measurements land within 0.1–0.5% of true distance, and stockpile volumes come in at 1–3% on well-defined, well-textured piles.

Those figures are RMSE measured at independent checkpoints withheld from processing. That distinction matters more than it sounds.

What happens without ground control, and why it is worse than the number suggests

Metres of horizontal shift is bad. The vertical situation is categorically worse, and three separate problems stack up.

GNSS vertical is inherently worse than horizontal — roughly one and a half to two times — because all your satellites are above the horizon.

DJI's altitude metadata is documented as unreliable. Agisoft publishes a support article and two scripts specifically to work around cases where the absolute altitude written into the EXIF is simply wrong.

Ellipsoid versus orthometric height. In this part of Arizona the geoid separation is roughly negative 25 to 27 metres. A pipeline that treats ellipsoid heights as orthometric puts the entire model about 85 feet out in Z. That is a bulk offset rather than noise, which is why one known elevation removes almost all of it — and why relative measurements survive it intact.

And then the part that is not an offset at all: peer-reviewed testing consistently finds bowl or dome shaped vertical distortion in nadir-only blocks flown without control. The surface is not just displaced, it is bent. The ISPRS study found the distortion was "effectively removed" once four corner control points were introduced — but until they are, even relative measurements across the block degrade.

Practical consequence: without ground control we will not deliver elevations, contours or cut and fill. We will deliver imagery and shape, and say so.

The thing that actually limits accuracy

It is not the drone.

Independent benchmarking on identical imagery found ground control set with consumer GPS produced roughly 8 cm accuracy, while the same flight with survey-grade GNSS control produced roughly 2 cm. A four-fold difference, from the receiver used to locate the targets rather than anything in the air. Pix4D states the ceiling directly: absolute accuracy cannot exceed the accuracy of the control.

So when we quote a job with stated accuracy, the ground control line is not padding. It is the line that makes the number true.

How many control points, and where

The research is clear that distribution matters roughly as much as count. Poorly distributed points give about twice the error of well distributed points at the same number.

• Even coverage including the interior, not just the perimeter. Corners-only is the classic mistake and it leaves the middle free to dome.
• Five points — four corners and a centre — is a functional minimum.
• Eight to twelve is the practical sweet spot for a 10 to 50 acre site.
• Returns saturate. Beyond roughly three to four points per hundred photos the improvement stops being meaningful.
• Vertical control at both the high and low ends of the site's relief.
• Always withhold three to five as checkpoints. A point used in the adjustment cannot validate the adjustment.

Where this sits against an actual survey

Honestly, and without flattery.

At its best, this produces data roughly two to four times less accurate horizontally than a conventional survey, and two to three times less accurate vertically — over a continuous surface rather than at discrete verified points. That is a genuinely useful engineering product and it is not a survey.

What this is good for, and what it is not

One more caveat worth stating: photogrammetry produces a surface model, not bare earth. In sparse Cochise County desert that is mostly favourable, but creosote, mesquite and acacia hold the surface up and need classifying out if you want true ground.

Local conditions

This is good photogrammetry country. High contrast and sparse vegetation give the software plenty of features to match, which is more than half the battle.

The failure modes here are specific: uniform caliche and bright sand are low-texture surfaces where matching degrades, and accuracy is always worst toward the edges of a block. Four to five thousand feet of density altitude shortens flights. And with no terrain-following in the flight app, rolling ground means either a third-party mission planner or accepting varying ground sample distance across the site.

What we would rather you asked

Not "is it survey grade." The useful question is "what decision are you making with this data, and what tolerance does that decision need?"

Balancing cut and fill on a building pad, sizing a stockpile for a royalty payment, working out where water goes on a parcel before you site a house, documenting a site before and after — all of that lives comfortably inside these numbers. Setting a fence on a property line does not, and no drone changes that.

Tell us the decision. We will tell you whether this reaches it, including when the answer is no.