Skip to main content
Hemyar Alharazi

Technical note · UAV

Demonstrating Accuracy in UAV Photogrammetry: Why Checkpoints, Not Control Points, Tell the Truth

Abstract

Accuracy figures in drone survey reports are often residuals on ground control points — numbers that describe how well the software fitted the constraints it was given, not how accurate the model is. This note sets out a checkpoint-based validation method, how to place and survey independent checkpoints, and how to report horizontal and vertical error so a client can judge whether a deliverable supports the decision it was bought for.

The problem with control-point residuals

A bundle adjustment is solved using ground control points (GCPs). Their residuals therefore measure internal consistency: how closely the solution honours points it was forced to honour. A project can report a 1 cm GCP RMSE and still carry several centimetres of error everywhere else — especially away from the control, at the edges of the block, and in elevation.

Independent checkpoints

A checkpoint is surveyed exactly like a GCP but withheld from processing. Because the solution never saw it, the difference between its surveyed and modelled coordinates is an honest sample of error.

Practical rules:

  • Survey checkpoints with a method at least three times more accurate than the accuracy you intend to claim.
  • Distribute them across the whole area, including edges and changes in terrain — not clustered near the control.
  • Use enough of them for the statistics to mean something; a handful is an anecdote, not a test.
  • Mark them so they are unambiguous in the imagery at the planned ground sampling distance.

Reporting error

Report horizontal and vertical error separately, as root-mean-square error:

RMSEz=1n∑i=1n(zmodel,i−zsurvey,i)2RMSE_z = \sqrt{\frac{1}{n}\sum_{i=1}^{n}(z_{model,i} - z_{survey,i})^2}

and the equivalent for easting and northing combined. State the number of checkpoints, how they were surveyed, and where they sat. If a standard is being claimed (for example an ASPRS accuracy class), name it and show the calculation it requires.

RTK and PPK do not remove the need

Direct georeferencing with RTK or PPK can reduce the number of GCPs needed, but it does not validate itself. Camera calibration, lever-arm offsets and timing errors all show up as systematic error that only an independent checkpoint reveals. A small number of checkpoints is still the cheapest insurance on the project.

What a client should ask for

  1. Checkpoint residuals, not only GCP residuals.
  2. Horizontal and vertical RMSE, reported separately.
  3. A map of where the checkpoints were.
  4. The survey method used for the checkpoints themselves.

A deliverable that answers those four points can be trusted for the decision it supports; one that cannot is a picture, however sharp.

Share