The disadvantages of photogrammetry: where it fails, and why

Photogrammetry’s strengths are well advertised. Its limits are better learned here than on a site where a stockpile comes out with a hole in it and the client is waiting. Each limitation below has a cause, and nearly every cause has a fix: a change to the flight, a target on the ground, a different camera, or the honest answer that another instrument should do this one.

It only models what the camera saw

The first and largest limit. A photograph records surfaces that faced the camera and were lit. Everything else does not exist in the model: the ground under a tree, the floor of a tunnel, the underside of a conveyor, the back of a stockpile against a wall. The software does not know what is missing and will happily stretch a surface across the gap.

The fix: fly when the leaves are off, where that is possible; add oblique and low photographs for faces and undersides; and for ground under canopy, use LiDAR, whose pulses reach the soil through gaps. For a model, mask the areas that were not seen rather than trusting the fill.

Surfaces it cannot match

The matcher needs texture: small, distinctive, stable patterns it can recognise from another angle. Some surfaces have none, or have texture that changes with the viewpoint, and on those the model is thin, lumpy, or invented.

Water

Reflects the sky and moves. No stable texture at all. Mask it; measure to the waterline.

Glass and polished metal

Reflections that move with the camera. The matcher follows the reflection, not the surface.

Blank surfaces

Fresh concrete, plaster, asphalt, snow, sand. Nothing to find. Targets, chalk, a cross-grid, obliques.

Very dark surfaces

Black rubber, wet coal, shadowed pits. Too little light reaches the sensor to show texture. Expose for them, or light them.

Repeating patterns

Tiles, gratings, brickwork. Features that match the wrong copy of themselves. Obliques help the geometry reject the mismatches.

Thin things

Wires, rails, rebar, fences. No texture on either side of an edge. They vanish; LiDAR keeps them.

The model can bend

A block of nadir photographs flown in parallel lines at one height is symmetrical in a way that lets a systematic error hide. The solver is also estimating the lens: how it bends straight lines towards the edge of the frame. A slight error in that bend and a slight curve in the ground produce the same pictures, and with nothing to say otherwise the solver splits the difference. The result is a model that domes up or bowls down towards its middle, by centimetres on a small site and by metres on a long, thin corridor.

The fix: break the symmetry. A few lines with the camera tilted 20–30° give the solver views that a curved world cannot satisfy; a second set of lines at right angles does the same; ground control points pin the edges to the truth; and flying a long corridor at two heights helps. A camera calibrated beforehand, and held fixed in the solution, removes the ambiguity at source.

The camera can distort it

Most small drone cameras read their sensor a row at a time, over a few tens of milliseconds. From a moving drone, each row sees the ground from a slightly different place, and every straight edge leans. At 10 m/s and a 30 ms readout that is 30 cm between the top and bottom of a photograph: more than ten pixels at a typical survey’s resolution. A study of three processing packages found that this distortion “needs to be modelled and cannot be eliminated by increasing the control points”.

The fix: a camera with a global shutter, or a rolling shutter with a short readout and software that models it (it needs the readout time); failing that, fly slower and higher, which reduces the shift per photograph. The same goes for focus and zoom: a lens that changed between lines is two cameras, and the solver has one lens model to fit them with. Lock focus at infinity and never zoom.

Height is the weak 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. On a nadir survey the vertical error runs two to three times the horizontal with good control, and far more without: the published test on the accuracy page found 23 cm of vertical error with RTK alone and nadir photographs, against 2.5 cm horizontal.

The fix: oblique photographs (which widened the angle enough to bring that 23 cm to 4 cm with no other change), more overlap, and ground control points at the edges and at the high and low ground. Vertical accuracy is bought with geometry and targets, not with megapixels.

It knows no scale by itself

Nothing in a photograph says how big anything is, or where it is. A block of photographs solved alone is a true shape at an unknown size, in no particular place, tilted however the solver found convenient. Measured in that model, a 10 m wall might be 10.3 m, and the whole site two metres from its design. This is not a defect of the method so much as a fact about photographs, but it catches people who expected a model to be a survey.

The fix: tie it to the world. RTK or PPK positions for the photographs, ground control points, or scale bars in close-range work. And check points held back from the solution, so the result is measured and not assumed.

Moving things and changing light

A haul truck that moved between photographs appears in several places at once, or as a smear, because the matcher is reconciling pictures of different scenes. People, vehicles, water, waving vegetation, and the drone’s own shadow all do this. Changing light does it more subtly: hard shadows move between lines, an automatic exposure brightens over the pit and darkens over the stockpile, and the same ground looks different in neighbouring photographs.

The fix: fly when the site is quiet, in even light; lock exposure and white balance; and accept that moving things are not surveyed, only recorded. For the moving things themselves, a position feed and a live map is the right tool, not a model.

Capture takes planning

The last limitation is the one that decides whether any of the others show up. A laser scanner measures what it is pointed at; a photogrammetry survey is decided by overlap, height, light, control and geometry before the first picture, and a poor plan cannot be processed into a good survey. Processing is also slow (hours for thousands of photographs) and needs a serious computer or a cloud service, so the result is not available in the field the way a scan is.

All of it in one table