C-DRONE GUIDE · 31 AUGUST 2026

Checking a Drone Photogrammetry Deliverable: Processing Report, Check Points, Tolerances

You receive an orthophoto, a point cloud, a digital terrain model and a PDF processing report. Somewhere in that PDF, a line reads « RMS : 0.018 m ». The contractor explains that the survey is therefore accurate to under two centimetres. That is wrong — not because he is lying, but because the figure describes how well the computation fitted its own ground control, not how accurate the result is. A survey can show a one-centimetre RMS and be twenty centimetres out in height in the middle of the site. The only figure that measures real accuracy is the deviation observed on check points excluded from the computation. Here is how to read a processing report, what to require before the flight, how to write a tolerance that holds up in a tender, and what an independent check costs.

Published on 31 August 2026, reviewed on 11 September 2026 — regulations in force as of September 2026.

Ground control points and check points: the costliest confusion

A drone photogrammetric survey rests on a bundle adjustment : the software simultaneously reconstructs the position of every image, the camera's internal parameters and a point cloud, then ties the whole thing to markers of known coordinates. Those markers are the ground control points (GCPs) : targets on the ground surveyed by GNSS or total station, whose coordinates enter the computation as constraints.

At the end of processing, the software displays the residual difference between the measured position of those control points and the position the adjusted model assigns them. This is the figure almost every report presents as « the accuracy of the survey ». Yet it is only a measure of fit : it says how well the model managed to pass through the points imposed on it. An adjustment can sit very close to its constraints while drifting substantially between them — that is in fact the typical behaviour of an over-parameterised model. Quoting an accuracy from that value amounts to grading a student only on the questions whose answers you handed them.

Check points (CPs) answer the other question. They are points surveyed on the ground to the same standard as the control points, but deliberately excluded from the computation. Once the model is built, their measured coordinates are compared with those the model gives them : the resulting deviation measures accuracy, because the model never saw those points. It is the only statistic that describes what the deliverable is worth where nobody placed a target.

This independence principle is no specialist's opinion : the ASPRS Positional Accuracy Standards for Digital Geospatial Data, in edition 2 (2023, revised version June 2024), make it an explicit requirement — the check points used for assessment must be independent of the control points used in the process, and their own accuracy must be at least twice as good as the accuracy expected of the tested product. The same framework retains RMSE as the sole accuracy measure, broken down into RMSEx, RMSEy and RMSEz, and recommends a minimum of 30 check points for a standard project.

French law says the same thing in its own vocabulary. The order of 16 September 2003 on precision classes applicable to categories of topographic work carried out by the State, local authorities and their public bodies — the text that replaced the 21 January 1980 order on tolerances — sets out in article 3 a safety coefficient C, defined as the ratio between the precision class of the points to be checked and that of the checking determinations, and requires that coefficient to be at least 2. In other words : you do not check a decimetric survey with decimetric measurements. Any French authority commissioning a survey is applying, often unknowingly, a text that describes exactly the acceptance procedure discussed here.

The gap between the two populations is far from anecdotal. A study by E. Sanz-Ablanedo, J. H. Chandler, J. R. Rodríguez-Pérez and C. Ordóñez, published in 2018 in Remote Sensing, used a particularly solid test case : over 1,200 hectares, more than 100 ground-surveyed points, more than 2,500 photographs, and 3,465 different combinations of control points tested in the adjustment, with accuracy assessed each time separately on the control points and on independent check points (see the study on Google Scholar). The authors show, with figures, how strongly accuracy depends on the number and distribution of control points — and therefore how blind to its own shortcomings a statistic computed on those points alone really is.

Keep the reading rule : a report that does not separate two populations of points supports no conclusion about accuracy. That does not necessarily make it a bad survey ; it makes it an unchecked survey, which is a different matter. On the principle of centimetric georeferencing and on what RTK or PPK do — and do not — bring to that chain, our guide to the RTK/PPK drone and centimetre accuracy covers the technical groundwork.

Reading the processing report: reprojection error, dome effect, systematic bias

The processing report almost always shows a second statistic alongside the RMS on control points: the mean reprojection error, expressed in pixels. It measures the gap between a point's position as computed by the algorithm in 3D space and its actual position on every photo where it appears. According to the documentation of Pix4D, one of the most widely used photogrammetry packages, this error should stay at or below one pixel; on camera optimisation, a relative gap between the initial internal parameters (from the manufacturer's spec sheet) and the parameters optimised by the computation above 5% deserves scrutiny, and beyond 20% signals a significant error — insufficient imaging geometry, poorly distributed control points, or an unsuitable camera model.

That statistic, however, says nothing about a more insidious flaw: the dome effect, a systematic dome-shaped deformation that affects models reconstructed from nadir-only (vertical) imagery, particularly over large flat areas. M. R. James and S. Robson, in a landmark study published in 2014 in Earth Surface Processes and Landforms, showed that this bias stems from imperfect self-calibration of the camera during the computation, and that it is effectively corrected by adding oblique imagery to the flight plan — a practice now common but still too often missing from tender specifications (see the study on Google Scholar). A processing report that looks flawless on paper — low RMS, sub-pixel reprojection error — can still hide a dome several centimetres high in the middle of the site if the flight plan was nadir-only: that is something to check before the flight, not after the deliverable.

Writing an enforceable tolerance: precision classes, check points, georeferencing

The order of 16 September 2003 mentioned above does more than set a safety coefficient: it gives project owners the vocabulary to write a tolerance that holds up in a tender. A precision class is expressed through a population of points and three cumulative conditions: a mean positional error below a given threshold, a maximum number of deviations exceeding that threshold (the "error template" of article 4), and an absolute maximum deviation no point may exceed. Article 6 further allows planimetry (x, y) and height (z) to be specified separately, which makes practical sense: a photogrammetric point cloud is almost always better in planimetry than in height, where a lack of texture or low vegetation degrades the reconstruction. A tender that demands "2 cm accuracy" without stating which axis, which population of points, and which error template applies sets no enforceable tolerance at all — it sets a slogan.

The number of check points directly affects how reliable that figure is. A study by P. Martínez-Carricondo and co-authors, published in 2018 in the International Journal of Applied Earth Observation and Geoinformation, systematically varied the number and layout of ground control points on the same site and measured the effect on the accuracy obtained at independent check points: below a certain density threshold, accuracy degrades sharply and unpredictably, particularly in height (see the study on Google Scholar). One question remains: the reference frame. A survey that is accurate in its own coordinate system can still be unusable if it is not tied to the right height datum — the RAF20 grid links RGF93 v2b and NGF-IGN69 across mainland France, while the RAC23 grid does the same for Corsica with NGF-IGN78; our guide to the coordinate system of a drone survey covers this link, essential for the negotiated tolerance to mean anything at all.

Frequently asked questions

Request a free quote

Put it into practice

• Drone surveying & photogrammetry: rates and cities covered from €800
• Drone security & surveillance: rates and cities covered from €600
• Surveying & photogrammetry in Bourg-en-Bresse Auvergne-Rhône-Alpes
• Surveying & photogrammetry in Nevers Bourgogne-Franche-Comté
• Surveying & photogrammetry in Paris Île-de-France

Also worth reading

• Olive orchards and drones: water stress and yield forecasting, price
• Prune d'Agen and drones: water stress and ripeness, price
• Asparagus and drones: spotting weeds on the ridge before emergence, price