How to fly a drone for photogrammetry: overlap, height and GSD

Photogrammetry is forgiving of almost everything except the flight. A good flight processes into a clean map on any software; a bad one does not process into a good map on any. This guide is the handful of settings that decide which you get, the reasons behind each, and a calculator for the numbers.

What the software needs from you

Photogrammetry finds the same point on the ground in several photos taken from different places and works out where it is in three dimensions from how its position shifts between them. Everything about a mapping flight follows from that sentence. Each point must appear in several photos (overlap). The photos must be sharp enough to recognise it (resolution, shutter speed, light). The camera's positions must be spread (a grid, not a hover). And the ground must look the same in every photo (nothing moving, no changing shadows).

1. Ground sample distance decides the height

Ground sample distance (GSD) is how much ground one pixel covers. It is set by the camera's sensor and lens and by the height above ground, and nothing else. The accuracy you can expect from the finished map is one to three times the GSD horizontally, and about twice that vertically, so choose the GSD from what the survey is for and the height follows. (What accuracy to expect.)

Height is measured above the ground under the drone, not above the take-off point. On a site with 40 m of relief, a flight set to 80 m above take-off is at 40 m over the high ground (fine GSD, more photos than planned) and 120 m over the low (coarser GSD than planned). Terrain following, which most flight apps offer from an elevation model, holds the GSD steady; without it, plan from the highest ground.

The same calculator has a page of its own, with the formula worked through: GSD calculator.

2. Overlap decides whether it processes

Front overlap is how much each photo shares with the next one along the line; side overlap is how much each line shares with the neighbouring line. A point needs to be in at least three photos to be placed, and more is better, especially on ground with little texture.

3. Camera settings

• Shutter speed fast enough that the drone's motion does not blur the pixel: at 8 m/s and 2.7 cm per pixel, 1/1000 s moves 0.3 of a pixel, which is fine; 1/250 s moves 1.2 pixels, which is soft. Set 1/1000 s or faster and let ISO and aperture follow.
• ISO as low as the shutter allows; noise is false texture that the matcher chases.
• Focus at infinity, locked. Autofocus hunting between frames makes images that do not agree.
• Exposure fixed for the flight when the light is steady, so adjoining photos match; auto when cloud is moving through, so none are blown out.
• Format: the highest-quality JPEG the camera offers. RAW is not needed for mapping and slows everything down.
• Nadir (camera straight down) for the main grid. Add an oblique pass at 20 to 30° around anything vertical.

4. Light, wind and time of day

The best light for photogrammetry is bright and overcast: no shadows, even exposure, every surface textured. Second best is a clear sky with the sun high, within two or three hours of noon, so shadows are short and do not move much during the flight. The worst is low sun, when long shadows hide half the ground and move between the first line and the last, and broken cloud, when the same ground is bright in one photo and dark in the next. Wind matters less than people think for the drone and more for the subject: moving grass and trees match badly, and a gust at the wrong moment smears a frame.

5. The flight plan

1. Extend the grid past the boundary by two lines and a few photos, so the area you care about is well inside the solved model. The edges of a model are always the worst part of it.
2. Fly lines across the wind rather than into and with it, so the ground speed is even.
3. Lines along the long axis of the site: fewer turns, fewer wasted frames.
4. Two grids at right angles (a crosshatch) when the site has tall features or you want the best possible elevation model. It doubles the photos; on an ordinary site it is not needed.
5. Battery swaps: resume where the plan stopped, within the same session of light. A flight finished the next morning is two flights.

6. Ground control, checkpoints, RTK and PPK

None of the above makes the map sit in the right place on the earth; it makes the map the right shape. Position comes from ground control points measured by a surveyor, or from an RTK or PPK receiver on the drone. If you are flying with PPK, remember the three things that are not in the photos: start the raw satellite logging before take-off, keep the base or reference station within about 20 km, and copy the log files off the drone with the photos. If you are flying with RTK, check the fix status in the log afterwards; a leg flown on a dropped link reverts to ordinary GPS. In either case lay three or four checkpoints and measure them: they are the only proof the flight landed where it says. (GCPs, RTK and PPK compared.)

What a good flight looks like on the ground

A 12 hectare (30 acre) earthworks site, flown at 80 m with a 20 MP 4/3 camera, 80% front and 70% side overlap, nadir, 8 m/s, at noon under high cloud: about 2.1 cm per pixel, 13 lines, 300 photos, 15 minutes in the air on one battery. Processed, it is a 2 cm orthomosaic and a 4 cm terrain model with no holes, and with RTK on the drone it sits on the design within 3 cm vertically at four checkpoints. Clarity's processing reports the checkpoint residuals, the fix rate and the base station on the survey itself, so the flight's quality travels with the map. (Drone image processing in Clarity.)