The Orthophoto That Leans
A UAV photogrammetric block covers a 12 km² flood-risk mapping project: 1,450 images flown at about 120 m above ground with a 24 MP camera, giving roughly 3 cm GSD, with 80% forward and 60% side overlap. Deliverables are an orthophoto and a DSM on a 10 cm grid, with a specification of horizontal RMSE ≤ 0.10 m and vertical RMSE ≤ 0.15 m. Six ground control points were used, surveyed with a handheld GNSS receiver. Checkpoint RMSE after bundle adjustment is 0.09 m horizontal but 0.42 m vertical. Roof outlines at the edges of the block are displaced by 0.3–1.2 m relative to the cadastral parcel layer, and building facades are visible. About 40% of the checkpoints fall under a riparian corridor of 8 m tall vegetation.
1. Explain the 0.42 m vertical RMSE and set out what you would change before re-flying or re-processing.
Read the worked answer
The horizontal result (0.09 m) and the vertical result (0.42 m) come from the same adjustment, so the problem is specific to height. Candidate causes, in order of likely magnitude:
- Checkpoint quality, not model quality. Six GCPs and checkpoints surveyed with a handheld GNSS receiver carry 2–5 m uncertainty. A checkpoint RMSE computed against such points is partly a measure of the checkpoints. Before blaming the model, re-survey the checkpoints by RTK or total station to 1–2 cm.
- Vegetation. 40% of checkpoints are under 8 m riparian canopy. A photogrammetric surface represents the canopy top, so a checkpoint on the ground under trees will misfit by metres regardless of how good the adjustment is. Checkpoints for a ground-surface specification must be on open ground; canopy points test something else entirely.
- GCP distribution and height control. With only six GCPs, vertical control is thin; a poor height on one or two GCPs propagates as a systematic tilt or bowl across the block.
- Block geometry. Cross strips and sufficient side overlap control the vertical bowl effect; check whether any were flown.
Changes
- Re-survey 12–20 GCPs and checkpoints with RTK, well distributed and including the block edges.
- Exclude canopy checkpoints from the ground-surface accuracy statistic and report them separately as a vegetation-penetration limitation.
- Add cross strips and enable self-calibration in the bundle adjustment.
- Re-process and re-report. Note that no amount of processing removes a canopy bias - that is a limitation of the sensor, not of the software.
2. A client complains that buildings lean and do not sit on their parcels. Quantify the cause and say what can and cannot be fixed.
Read the worked answer
Cause: relief displacement For an image taken at height H above ground, a point at height h and radial distance r from the nadir point is displaced by d = (r × h) / H.
Here H ≈ 120 m. With a 180 m-wide image footprint the half-width is r ≈ 90 m. A 12 m building at the block edge gives d = (90 × 12) / 120 = 9 m of raw relief displacement in the individual photographs.
Orthorectification against a DSM removes most of this, reducing it to the observed 0.3–1.2 m. The residual comes from:
- the DSM being a smoothed 10 cm grid that does not capture vertical walls;
- occlusion - surfaces hidden behind a building in every image cannot be recovered, so they smear radially;
- the DSM edge producing a step rather than a wall.
What can be fixed
- Produce a true orthophoto: build a building/DSM model with explicit walls and footprints, and correct occlusion by selecting, for each ground pixel, an image in which that pixel is actually visible.
- Tighten the DSM grid and improve the point-cloud classification of buildings.
What cannot be fixed
- Ground beneath a roof overhang or in a deep occluded courtyard is not imaged at all. No processing recovers it. That area must be annotated as no-data or surveyed by another method.
- A 10 cm orthophoto will never have the geometric fidelity of a 1:2,000 cadastral survey; it must not be used to resolve parcel boundaries.
3. The client intends to use the DSM directly as terrain input for a flood model. Assess this and state what you would certify.
Read the worked answer
Assessment A photogrammetric DSM is a surface model: it records the first return, which over vegetation is the canopy and over buildings is the roof. A hydraulic flood model needs a terrain (bare-earth) model. Using the DSM in the riparian corridor - 8 m of vegetation - would raise the modelled ground by several metres and divert or dam the flood plain in exactly the area the project cares about.
Additionally, the vertical specification of 0.15 m RMSE is currently not met (0.42 m measured), and the measured figure is itself contaminated by the canopy checkpoints described above.
What you do
- Do not certify the DSM as meeting the vertical specification. Report the achieved accuracy honestly, and state which checkpoints were in the open and which were under canopy.
- Classify and filter the point cloud to derive a DTM, and quantify the DTM against open-ground checkpoints. In dense riparian canopy, photogrammetry may not penetrate to the ground at all; if the DTM error in that corridor exceeds the model's sensitivity, recommend a LiDAR or ground-survey supplement for that strip.
- Deliver both products with explicit metadata: DSM (surface, includes vegetation and buildings) and DTM (bare earth, with stated limitations and the corridor flagged).
- Advise the client in writing that flood modelling must use the DTM, and that the DSM is suitable only for roughness/obstruction or visualisation purposes.