What accuracy can you expect from deviation detection?
Customers often ask us for a single accuracy figure, usually because a supplier specification quotes a tolerance such as 1 cm. There is no single figure, because three separate things determine whether a deviation we report is real. They are frequently confused with one another, so this article explains all three.
1. The accuracy of the dataset
How closely the processed model matches the real world. This is the layer you control through how the site is flown, and with RTK and correctly placed and surveyed ground control points, 2-5 cm is realistic.
Two things are worth knowing. The number of ground control points is not what determines accuracy: how they are placed across the site and how accurately they are surveyed in matters far more. And vertical accuracy is weaker than horizontal, typically by a factor of 1.5 to 2.
> RTK for relative and absolute arccuracy covers this layer in detail.
2. The accuracy of the detection
Once the model exists, every component still has to be identified within it. For a pile, that means pinpointing where it was actually driven into the ground.
This is normally the layer that limits what we can tell you, not the dataset. Two things drive it:
How clearly the component appears in the orthomosaic. One with poor contrast, or with no visible footprint or shadow, is harder to pin down.
How consistently your design aligns with the as-built positions. If the design sits 10 cm off in one part of the site and 30 cm off in another, the comparison inherits that inconsistency.
Piles are the hardest component to identify precisely: they are small, cast little shadow, and their exact centre is a matter of judgement even in good imagery. Trench and road edges read more clearly. For trenches, though, an orthomosaic shows the open excavation at the surface, whose width is shaped by soil battering and spoil heaps rather than by the designed width, so depth is the more meaningful dimension to compare.
3. The threshold at which we flag a deviation
The threshold a component has to exceed before we flag it sits deliberately above the detection accuracy. Set any lower, and normal variation in identifying a component would produce false deviations on every flight.
A threshold is therefore not a statement of how accurately we can measure. It is the point beyond which a difference from the design is real rather than an artifact.
The current thresholds per component type are listed in How to identify and manage deviations in construction, and can be adjusted for your site. Racking and structures are tracked for progress today but do not yet have a deviation check.
A note on trench depth
We measure depth as the difference between the trench floor and the ground immediately beside it, on the same cross-section. Both readings come from the same place in the model, so any overall vertical offset affects them equally and cancels out. Depth is a relative measurement, and does not inherit the absolute vertical accuracy of the dataset.
What limits it is whether the floor is visible at all: a narrow or deep trench can be partly occluded from above, and where we cannot see the floor we cannot measure the depth. Where we do measure a stretch, we report the shallowest section along it rather than an average, so a single shallow stretch is not hidden by the rest being dug to specification. The depth we report errs on the conservative side.