For a map intended to support earthworks quantities, repeat surveys or asset measurements, the question is not whether the receiver showed "FIX." It is whether the delivered coordinates have been compared with survey measurements that the reconstruction did not use.
An accurate camera position is only one input
RTK or post-processed kinematic (PPK) GNSS can improve the georeferencing of camera stations. It does not directly measure the location of every point on a wall, road or field. Structure-from-motion processing matches features across overlapping photographs and adjusts camera poses and camera calibration parameters to reconcile those observations.
A systematic camera timing offset, the wrong coordinate reference system, weak image geometry or unstable calibration can affect the reconstructed surface even when individual camera coordinates are tightly estimated. The US Geological Survey's guidelines for UAS imagery calibration treat flight design, ground control, tie-point quality, camera calibration and independent checks as linked parts of geometric quality assurance. Its analysis also notes that vertical measurement can be less constrained than horizontal position.
Ground sampling distance (GSD) describes the image detail projected on the ground. It is not a measurement of absolute map accuracy. A two-centimeter GSD does not certify two-centimeter coordinate accuracy. Pix4D's image-acquisition guidance separates image detail from the precision of georeferencing and recommends well-distributed, surveyed control where required.
Why a good bundle adjustment can still hide a bad map
A processing report may show small image reprojection residuals and low ground-control residuals. Both can be useful diagnostics, but neither independently establishes accuracy away from the observations that constrain the adjustment.
The reason is structural. Ground control points (GCPs) enter the model as known coordinates. The bundle adjustment is expressly trying to fit those coordinates, alongside the image measurements. A small residual at one of its inputs says that the model fits that input. It does not reveal the error between those locations, where the final surface is used.
Likewise, a low reprojection error establishes that projected features and measured image positions agree within the adopted model. A biased lens model, repetitive texture or insufficient cross-strip geometry can still yield a surface with unwanted vertical curvature. Pix4D's quality-report documentation discusses bent reconstructions, camera-calibration problems, ground-control misidentification and the increasing uncertainty away from well-placed control.
These problems matter differently by output. A visually convincing orthomosaic may be adequate for documenting site conditions yet unsuitable for calculating small elevation changes. A model that looks good at roof level may not meet a vertical tolerance on a road surface.
Control points improve the solution; checkpoints test it
A checkpoint is a surveyed, photograph-identifiable point deliberately withheld from the georeferencing adjustment. The photogrammetry software estimates its coordinates from images, then compares those coordinates with the independent survey values. The difference is an external test of the result at that location.
Pix4D explicitly distinguishes GCPs used for georeferencing from checkpoints used for quality assessment. Its PIX4Dmatic documentation says checkpoints are not used to georeference the project. The distinction disappears if a supposed checkpoint has also influenced the adjustment as a control constraint.
Independence has a second layer. A checkpoint measured from the same poorly configured GNSS base, wrong vertical datum or common coordinate conversion can share systematic errors with the camera positions and GCPs. Withholding it from software optimization does not automatically make its survey reference error-free. An operator should document the base station, datum, units, correction source, monument/target identification and survey method.
Distribution matters more than a neat point count
Consider a long road corridor whose only surveyed targets cluster around the takeoff area. The adjustment may agree with those targets yet give poor evidence about accuracy at the far end. A second problem arises when all checks are on flat ground but the deliverable includes embankments, roofs or cut faces: the checkpoints fail to sample the vertical geometry of interest.
For a defensible project, identify the desired accuracy and product first, then distribute surveyed targets over the horizontal extent and relevant elevation range. Put independent checks in the regions where a wrong answer would matter. Check both interior and edge behavior where possible. Include targets that are unambiguous in multiple images, and ensure that targets are surveyed at the physical point being marked in the software.
There is no universal magic number for every site. Pix4D's tie-point guide recommends approximately five to ten GCPs for many projects, with additional surveyed points allocated to checkpoints. That is general software guidance, not a substitute for an accuracy specification or a sampling design for a large corridor. The USGS guidelines discuss repeat adjustments using different control/checkpoint partitions when a full independent network is impractical; that approach exposes sensitivity to particular control choices, although it does not remove common survey biases.
Cross-strip or multi-direction imagery can strengthen self-calibration by viewing the same stable features from different camera geometries. The USGS report discusses orthogonal patterns for this reason. It is often more useful to improve camera geometry and target placement than simply to add more nearly identical nadir images.
Read the error table as a spatial diagnosis
For every checkpoint, retain the differences between reconstructed and surveyed eastings, northings and heights. Report their mean errors, spread and root mean square errors (RMSE), separately for horizontal and vertical components, using a stated coordinate system and unit. Pix4D defines its point-error statistics in terms of these coordinate differences.
A mean vertical offset can indicate a systematic datum, base-height or model issue; a large spread without a consistent direction suggests a different investigation. Plot residual vectors or label them on the site plan. A single overall RMSE can conceal a pattern in which the center of a survey meets tolerance while the boundaries do not.
The project should state which accuracy criterion it is applying rather than attach an arbitrary "survey grade" label. The ASPRS Positional Accuracy Standards, Edition 2, Version 2 (2024) provide an industry framework, but the required class, verification method and reporting obligations depend on the deliverable and contract. Do not mix criteria from different standard editions or quote a confidence figure without the standard's method.
A checkpoint residual is also the sum of uncertainty in the map and the reference survey. If the required tolerance approaches the uncertainty of the check-point survey itself, a tighter survey method or an explicitly bounded statement of uncertainty is needed. USGS discusses this limitation and the option of using total-station measurements where GNSS-derived control is not precise enough.
What an operator should deliver with the map
An auditable map package needs more than an orthoimage, digital surface model and a screenshot of the RTK indicator. It should retain:
- The image set or traceable flight metadata, camera model and processing version, including any rolling-shutter or camera-calibration settings relevant to the result.
- The horizontal and vertical reference systems, base/correction source, GCP measurements and the survey uncertainty attached to them.
- A clear distinction between the points constrained in the adjustment and the withheld checkpoints.
- Checkpoint coordinate differences and summary statistics, plus a plot showing where the checks sit relative to the deliverable.
- Exceptions: missing coverage, weak texture, variable-height surfaces, vegetation, water, extrapolated edges or areas with no independent checks.
Those records let a later user judge whether a dataset is fit for a particular measurement. They also make a repeated survey comparable with an earlier one. A new flight processed with a changed datum, camera model or checkpoint policy can produce a difference surface that describes a workflow change rather than a change on the ground.
The practical boundary is straightforward: RTK is a useful way to constrain camera positions. Independent checkpoints, measured and reported to a defensible specification, are how an operator tests the final map. A fixed GNSS solution and an attractive reconstruction are not replacements for that test.