How accurate is photogrammetry? Real figures and what decides them

“How accurate is photogrammetry?” has an honest answer, but it is not a number. It is a short list of things that decide the number, and once you know the list you can read any survey’s accuracy off its flight plan and its control. This page is that list, with the published figures that back it up and the standard the figures are measured against.

Two kinds of accuracy

Nearly every disappointment with photogrammetry is a relative-accuracy expectation applied to an absolute-accuracy question: a model that looked perfect and measured volumes beautifully turned out to be a metre out against the engineer’s design, because nothing tied it to the ground.

What decides it

Four things, in roughly this order of importance.

The pixel

Ground sample distance. Nothing is measured finer than a pixel, and the usual result is one to two pixels horizontally. Halve the flying height, halve the error.

The control

Targets on the ground, positioned photographs (RTK/PPK), or both. Decides absolute accuracy entirely, and vertical accuracy almost entirely.

The geometry

Overlap, oblique photographs, cross lines. Decides how rigid the block is and whether it bends: the height error at the middle of a long nadir-only strip is geometry, not pixels.

The camera

A stable lens, a global or well-modelled rolling shutter, sharp exposure. A camera that changed focus between lines adds error no control removes.

The animation shows why height is the harder axis. A point’s position across the picture is pinned by a single line of sight; its height depends on the angle between two lines of sight, and from a drone looking straight down that angle is small. Wider overlap and oblique photographs widen it. This is why vertical error runs two to three times the horizontal on a nadir survey, and why the first fix for poor heights is not a better camera but a different flight.

The budget, interactively

Published figures

The clearest recent test is from the ISPRS annals of 2024: a 20 MP RTK drone flown at 35 m over a square, checked against 71 points surveyed with a total station, with and without targets, with and without oblique photographs. The figures below are its check-point RMSE.

Three lessons sit in that table. Horizontal accuracy was between 1.4 and 2.9 cm whatever was done: close to the pixel, as the rule says. Vertical accuracy swung from 23 cm to under 1 cm on the same site with the same camera, decided entirely by control and geometry. And oblique photographs alone took the vertical error from 23 cm to 4 cm with no targets at all, which is the cheapest improvement in surveying.

1.5–2.9 cm

Horizontal, every configuration tested

23 cm

Vertical: nadir only, RTK, no targets

4.3 cm

Vertical: oblique added, still no targets

0.9 cm

Vertical: RTK, obliques and nine targets

Sensor makers’ own figures line up with this: a current 45 MP survey camera on an RTK drone is quoted at about 3 cm horizontal and 5 cm vertical without targets, and drone LiDAR at 5 cm vertical and 10 cm horizontal; both sides of the comparison are on the LiDAR page.

The standard it is tested against

In North America and widely elsewhere, accuracy is reported to the ASPRS Positional Accuracy Standards for Digital Geospatial Data, whose second edition (2023) changed the ground rules in ways worth knowing.

One figure

Accuracy is the RMSE against check points, horizontal and vertical, with a new 3D figure for point clouds. The old 95% confidence statement is gone.

No tie to the sensor

Classes are no longer set by pixel size, map scale or scanning resolution. Any sensor can meet any class if the check points say so.

Check points

At least 30 for horizontal and vertical testing, about twice as accurate as the product, and their own error is now counted in the result.

Vegetation

Vertical accuracy is tested separately on open ground (must meet the class) and under vegetation (reported as found, no pass or fail).

What to ask for on a report

A processing report says more than the number on the cover. The things to read, in order:

Photographs aligned

All of them, or an explanation. A whole line missing is a hole in the block.

Reprojection error

In pixels. A few tenths is good; over one, something in the camera model or the shutter is wrong.

Control residuals

The fit at each target after the solution. One bad target is usually a mis-click or a mis-survey, and shows here.

Check-point RMSE

Against points the software was not allowed to fit. The only accuracy figure to quote.

For how to fly so the report reads well, the 101; for the failures that put error into a survey and how each is cured, the limitations page; and for what accuracy means in the measurements a site actually needs, the drone survey accuracy guide. What those centimetres mean for a stockpile number is on the stockpile volumes page.