Planning a Photogrammetry Flight That Meets Its Accuracy Target

Start from the deliverable, not the drone

The first question is not what altitude to fly. It is: what decision will this data support, and what error would change that decision?

A stockpile volume for monthly reconciliation might tolerate 5 cm. An as-built comparison against design tolerances might need 2 cm. A visual condition survey might need no absolute accuracy at all. Each implies a different flight, a different amount of ground control, and a different cost.

Ground sample distance

GSD is the ground distance represented by one pixel:

GSD = (sensor pixel size × flight height) / focal length

For a given camera, GSD scales linearly with altitude. A useful rule of thumb is that final horizontal accuracy lands around 1–3 × GSD with good control, and vertical accuracy around 2–4 × GSD.

So if the deliverable requires 2 cm vertical accuracy, a 2 cm GSD is already marginal. Plan for roughly 1 cm and fly lower — accepting a longer flight and more images — rather than discovering the shortfall after processing.

Overlap

Photogrammetry requires every point to appear in several images taken from different positions. Sensible starting values:

• 75–80% frontal, 65–70% side for open terrain with good texture
• 85% / 75% for vegetation, water margins, or uniform surfaces such as fresh asphalt
• 90%+ for dense vegetation, or use oblique imagery instead

Under-overlapping is the most common cause of holes and warped reconstructions, and it cannot be repaired in processing. Flying twice costs far less than mobilising twice.

Ground control and check points

Ground control points are surveyed markers used to constrain the reconstruction. Without them, the model is georeferenced only by onboard GNSS, which for a non-RTK aircraft may be metres out in absolute terms — while looking entirely self-consistent.

Distribute at least five GCPs: near each corner and one central, with additional points where the site has significant relief. Vertical error grows fastest away from control, so elevation extremes deserve their own points.

Check points are surveyed but deliberately withheld from the solution, then used to measure error independently. A survey reporting only GCP residuals is reporting how well it fitted its own constraints, which is a much easier question than how accurate it is. Always keep some points back.

RTK and PPK reduce dependence on GCPs but do not remove the need for check points, because they validate a different part of the chain.

Execution details that decide quality

• Fly in consistent light. Moving cloud shadows across a site degrade image matching considerably.
• Watch shutter speed. Motion blur at flight speed is common and irrecoverable. Keep exposure short enough that ground motion during the exposure is well under one GSD.
• Cross-hatch for detail. A second perpendicular grid, or oblique passes, dramatically improves reconstruction of vertical faces, building edges and anything with an undercut.
• Avoid low sun angles unless shadow relief is specifically wanted; long shadows hide detail and confuse matching.
• Calibrate the camera or allow self-calibration with enough geometry to solve it — a fixed focus lens is easier to model than a zoom left to its own devices.

Report what you actually achieved

A defensible deliverable states GSD, overlap, control configuration, processing software and settings, coordinate reference system, and the check point residuals.

That last number is what a reviewer will ask for, and having it is what turns a picture into evidence.