Visible-light imaging and photogrammetry: millimetres per pixel decide what you can inspect

A visible-light camera records surface colour and shape: cracks, spalling and exposed rebar, rust, coating loss, water stains, efflorescence, missing bolts, deformed members. Most defects first show at the surface, which makes the camera the most widely useful inspection sensor.

Once you understand the path from lens to sensor to file, inspection camera spec sheets become much easier to read. Understanding the camera as a chain of parts (Airffic Aerial) is the shortcut.

GSD (ground sampling distance) is how many millimetres of the surface one pixel covers. The camera is a pair of similar triangles: GSD = sensor width × stand-off distance ÷ (focal length × image width in pixels). Halve the distance and GSD halves; double the focal length and GSD halves at the same distance.

For example, a 1-inch-class sensor (13.2 mm wide), 8.8 mm focal length and 5,472 pixels across gives the values below. These are geometric values; lens softness and focus make real images somewhat worse.

A crack narrower than one pixel may still appear as a faint line if it is darker than its surroundings. But measuring width is different: keep GSD comfortably finer than the width you need. Where you cannot get close, use a longer lens — but telephoto magnifies vibration and wind sway, so it goes hand in hand with blur control. Try your own camera in the stand-off and resolution calculator.

Drones shoot while moving, so motion blur appears. Blur in pixels = ground speed × exposure time ÷ GSD. At 2 m/s and 1/500 s the aircraft moves 4 mm; with a 1.4 mm GSD that is about three pixels of blur, enough to smear a fine crack away. Aim to keep it within one pixel.

• Slow down or stop: for close-ups, hover and shoot one frame at a time.
• Faster shutter: costs light, so raise ISO or add lighting — a trade-off against noise.
• Mechanical shutter: an electronic (rolling) shutter distorts moving shots and degrades photogrammetry.
• Gimbal and wind: gusts shake the aircraft, and the longer the lens the worse it shows. Re-plan distance and focal length on windy days. Wind & turbulence

Light decides quality too. Soffits and north faces are dark, and mixed sun and shade blows out highlights or crushes shadows. Overcast days with soft shadows suit inspection best. In the dark, use onboard lighting and lock exposure and white balance so every flight matches. Glare on wet or painted surfaces is usually fixed fastest by changing the angle. Keeping RAW files lets you lift shadows later. Darkness

Filmmakers obsess over shutter speed, frame rate and rolling-shutter distortion. Learning how to set shutter and fps from the aerial filmmaking side (Airffic Aerial) makes it easier to tell why an inspection photo is blurred.

Photogrammetry solves camera positions and object shape at the same time from overlapping photos. Today’s mainstream combines SfM (Structure from Motion) with MVS (multi-view stereo):

The basic rule is overlap — commonly 70–80% forward and 60–70% sideways. Purely nadir blocks can let lens distortion go poorly estimated, bowing the whole model (“doming”). Mixing in oblique images helps prevent it.

Running SfM yourself is the quickest way to see what goes wrong with too little overlap. Reconstructing shape from aerial photos with OpenCV and COLMAP (Airffic Aerial) is a good place to start.

Photogrammetry products serve different purposes. An orthomosaic removes perspective so the stitched image looks straight down (or straight at a wall); its uniform scale lets you measure lengths and areas. A 3D model (point cloud or mesh) keeps the shape itself.

You hear “digital twin” a lot, but a 3D model alone is not a twin. A twin useful for inspection links member IDs, past inspections, repairs and sensor data to the model and keeps being updated as reality changes. The AI and data management chapter covers how to connect them.

There are two kinds of accuracy: relative (correct lengths and shapes inside the model) and absolute (how well it matches map or drawing coordinates). Measuring a crack’s length needs only relative accuracy; overlaying with last year or placing on drawings needs absolute accuracy.

• Ground control points (GCPs): targets with known coordinates that lock the model to reality.
• Check points: points held out of the solution and used only to verify accuracy. Without them you cannot claim an accuracy.
• RTK / PPK: precise aircraft positions reduce the GCPs needed — but keep some check points.
• Scale bars: under decks or indoors without GNSS, include an object of known length to confirm scale. GNSS-denied
• Rule of thumb: in good conditions horizontal error is often said to stay within a few GSD, vertical usually worse. Decide the required accuracy from the inspection purpose first.

In Japan the Geospatial Information Authority publishes manuals for UAV use in public surveying; in the US, ASPRS defines positional accuracy standards for geospatial data. Borrowing the same idea for inspection — prove error with check points — is the safe path.

If inspection imagery is delivered as survey output, appearance and positional accuracy have to be reported separately. See separating appearance from accuracy in surveying and point clouds (Airffic Flight Institute).

• Below the surface: delamination, internal cracks and steel thickness loss are invisible.
• Beyond dirt and shadow: mud, moss, droppings and shadows hide defects — and mimic them.
• Featureless surfaces: new concrete, uniform paint, glass and water are hard to reconstruct.
• Darkness and time of day: unlit spaces, low evening sun and night degrade images badly.
• Judgement: photos are raw material. Interpreting defects is the engineer’s job.

By 2035, each structure will have a fixed capture plan: the drone shoots from the same poses, the same distances and similar light every time. GSD and overlap are checked automatically before take-off, and missing images are re-shot on the spot. Orthomosaics and 3D models are tied to member IDs, and only the differences reach the engineer. This is the editors’ view — but most of the technology is already in the field.

What changes most is not the hardware but the planning: from “fly and shoot” to “fly backwards from the pixel size you need”. When that is normal, inspection by drone has truly arrived.