Headline accuracy numbers from sensor manufacturers — ±10 mm, ±15 mm — describe what the sensor can do under perfect conditions. They don't describe what most captures achieve, or how to tell the difference. This is the engineering version.
If you've read three drone LiDAR brochures you've seen three different accuracy numbers, all confidently quoted, all technically correct, all referring to different things. The honest answer to how accurate is this stuff? requires unpacking what "accuracy" actually means in practice, where the error budget lives, and how to validate any claim against a number a third party can verify.
This article is the engineering answer rather than the marketing one. It applies whether you're scoping a project, reviewing a vendor's QA pack, or trying to understand why two captures of the same site report different residuals.
For drone LiDAR captured with engineering rigour in Australian conditions in 2026:
The numbers come from RMSE against independent ground control checkpoints — the only defensible measure of realised accuracy.
Three words are routinely conflated:
Accuracy is how close your measurement is to the true value. If the ground at a point is at AHD 100.000 m and the cloud reports 100.020 m, accuracy at that point is 20 mm.
Precision is how repeatable measurements are. If five separate captures of the same point all return 100.020 m, the capture is precise — and consistently wrong by 20 mm.
Resolution is how finely the cloud samples the surface — points per square metre. A 200 pt/m² cloud is higher-resolution than a 40 pt/m² cloud, but resolution and accuracy are separate variables.
A vendor that publishes "accuracy" without specifying which of the three they mean is — at best — being imprecise.
Realised accuracy at the surface is the root-sum-square of error contributions from every stage of the capture and processing pipeline:
| Source | Typical contribution | | ------------------------------- | -------------------- | | Laser ranging accuracy | 3–10 mm | | GNSS + IMU trajectory (post-PPK)| 10–20 mm | | Boresight calibration | 3–5 mm | | Strip alignment / overlap | 5–10 mm | | Ground classification + surface | 5–15 mm (DTM only) | | Control tie / datum residual | 3–8 mm |
Realistic combined accuracy on a well-run capture: 15–25 mm vertical RMSE on bare ground. The numbers add geometrically (RSS) not linearly, so trimming any single contributor has a smaller-than- expected effect on the final number — but the inverse is also true: a poorly-controlled trajectory or skipped PPK can push the whole result much wider.
The single biggest source of confusion in this space is sensor manufacturers quoting ranging accuracy (the laser's intrinsic ability to measure distance) as if it were system accuracy (what the geo-referenced point cloud delivers at the surface).
A sensor with 5 mm ranging accuracy is genuinely impressive. It's also one of six error sources in the table above. The geo-referenced point's position depends on knowing:
Any error in any of these compounds. A premium sensor on a sloppy capture routinely loses to a mid-range sensor on a rigorous capture by 30–50 mm. The same sensor in different hands produces wildly different results because the operator owns most of the error budget.
If you want to move the realised accuracy of a capture, control rigour is where the biggest lever sits.
The cloud is georeferenced through ground control points — surveyed marks tied to the Australian Height Datum (AHD) and the appropriate MGA zone. The denser and better-distributed the control, the better the cloud's tie to absolute datum. Sparse or one-sided control introduces systematic bias the cloud carries through every downstream deliverable.
For engineering-grade work, control density of one well-distributed point per 5–10 hectares is a reasonable target. For planning-grade work, a few perimeter points suffice. For PPK-only work without any ground control, the cloud is internally consistent but has no defensible absolute tie — appropriate for visualisation, problematic for engineering.
Every engineering-grade capture should be post-processed against a logged base station — PPK (Post-Processed Kinematic) — rather than relying on real-time RTK. The reasons:
A capture quoting accuracy without specifying PPK is one of two things: either using a different acronym for the same workflow (PPP, etc.), or relying on RTK and exposed to silent accuracy degradation on any link drop.
The most important thing on the QA pack is the residuals at independent control checkpoints — surveyed marks the LiDAR captured but that were NOT used in the calibration. The distinction is critical: marks used in calibration are guaranteed to fit the solution; marks NOT used in calibration measure how well the solution fits reality everywhere else.
A capture that reports residuals only at control marks used in calibration is reporting how well the maths converged — not how accurate the data is. Always ask which marks were used and which were withheld.
Downward-looking LiDAR achieves better vertical accuracy than horizontal accuracy for a simple geometric reason: small angular errors in the IMU project minimally into the vertical at typical flight altitudes (the laser is firing nearly straight down) but multiply into horizontal displacement (the further the pulse travels horizontally before hitting ground, the more leverage the angular error has).
Typical numbers:
The implication for design work: vertical-driven deliverables (DTM, cut/fill, contours, drainage) sit on tight tolerance. Horizontal-driven deliverables (asset positions, edge locations, CAD linework) need slightly looser tolerance assumptions, especially for features without a sharp edge.
The accuracy bands above assume bare or sparsely-vegetated ground. Under vegetation the realised accuracy degrades:
| Cover | Vertical RMSE typical | | ---------------------- | --------------------- | | Bare / sparse | 15–30 mm | | Moderate (suburban) | 25–50 mm | | Dense canopy | 50–80 mm |
Three reasons for the degradation. First, fewer pulses reach the ground, so the ground point density drops and the surface is interpolated across larger gaps. Second, classification can mistakenly include low vegetation as ground (or exclude genuine ground returns as vegetation) on edge cases. Third, the surveyed control checkpoints used for validation are themselves often near edges where canopy effects are highest.
A reputable vendor reports residuals separately by cover class so you can see which parts of the cloud are sitting on tighter tolerance than others. A vendor that doesn't is hiding the variance in a single headline number.
The most common patterns:
1. Sensor accuracy quoted as system accuracy. Brochure says ±10 mm; the sensor manufacturer's datasheet says ±5 mm ranging, plus GNSS/IMU/control/boresight. The 10 mm "system accuracy" claim is either a marketing number or measured under conditions you won't encounter in production.
2. Accuracy quoted on hard reference targets, not real terrain. A black-and-white survey target on a runway gives the sensor an ideal hit pattern. Real terrain — grass, gravel, asphalt with paint wear — doesn't. Always ask whether the residuals were measured on real surface or calibration targets.
3. Accuracy quoted without specifying point density. A 50 pt/m² cloud will produce more accurate surfaces than a 5 pt/m² cloud from the same sensor — sometimes by 2x. The number means nothing without the density it was achieved at.
4. RMSE quoted without the distribution. RMSE is a single number that hides the tails of the distribution. A capture with RMSE 20 mm and a maximum residual of 40 mm is reasonable. A capture with RMSE 20 mm and a maximum residual of 200 mm has localised problems the headline number doesn't surface. Always ask for the histogram or at least the worst-case.
5. "Internal consistency" quoted as accuracy. A capture can be internally consistent (overlapping flight strips agree with each other) while being absolutely wrong by 100 mm because the entire solution drifted on the GNSS. Internal consistency is necessary but not sufficient — only checkpoint residuals against AHD-tied marks prove absolute accuracy.
Five questions that cut through marketing and surface what's actually been measured:
What's the RMSE at independent ground control checkpoints? Ground control marks used in calibration don't count. Independent means withheld from the solution.
What's the maximum residual, not just RMSE? RMSE hides the tails. The 95th percentile residual or the absolute worst case shows the localised quality.
What surface were the residuals measured on? Pavement / grass / canopy edge — different surfaces give different numbers.
What was the point density at the checkpoint locations? Sparse density gives interpolated checkpoint heights, which is a different measurement than direct hit.
Was the trajectory PPK-corrected against a logged base? If the answer is "RTK was fine" or "we used PPP" without a base log, the accuracy ceiling is significantly lower than PPK can achieve.
Any reputable vendor will answer all five without hesitation. A vendor that bristles, deflects or quotes sensor spec instead of realised numbers is signalling something about the workflow you should take seriously.
The realistic engineering-grade accuracy workflow looks like this:
This is the workflow that produces accuracy numbers a third party can verify. It's not the only workflow that exists in the market, but it's the one that survives an independent audit. Every project should land with a QA report along these lines — without one, the accuracy claim is just a number.
Tell us the deliverable tolerance, control situation and surface conditions. We'll scope the capture to hit the number, with the QA workflow to prove it.
Every stage from site planning through QA to deliverable handover — including where the accuracy lives in the workflow.
Accuracy table comparing drone LiDAR, drone photogrammetry, traditional ground survey and helicopter LiDAR side-by-side.