Comparing a Drone As-Built Against the Design Surface

An earthworks job reaches the point where somebody wants to know whether the ground is built to the design. There is a design surface from the engineer and there is a drone as-built from last week. Subtracting one from the other is a few seconds of processing. The reason these comparisons go wrong is almost never the drone, and almost always that the two surfaces were never in the same frame to begin with.

This is the part of the workflow that generates the awkward meeting: a difference map showing the whole pad 180 mm low, an argument about whether the contractor underbuilt, and a week later the discovery that the design was on a different vertical datum. Worth understanding before the map is in front of anybody.

What the comparison actually produces

A surface-to-surface comparison takes two elevation models covering the same ground and reports the difference between them at every point. From that single operation you get the three things people actually want:

• An elevation difference map. A coloured raster showing, point by point, how far the built ground sits above or below design. This is the deliverable people mean when they say cut and fill map.
• Volumes to grade. How much material has to move to bring the as-built to design, split into cut and fill. Not the same number as net, and the net figure on its own hides a site that is 2000 cubic metres high in one corner and 2000 low in another.
• Conformance against tolerance. The area that falls inside the specified tolerance band and the area that does not, which is the question a contract usually turns on.

Note what is not in that list: a boundary, a legal elevation, or a certified quantity. Those are separate questions, and the last one is covered in the article on who can sign off drone survey data.

The two surfaces have to be in the same frame

This is the single largest source of wasted time on as-built comparisons, and it is worth being pedantic about because the failure is silent. A frame mismatch does not throw an error. It produces a clean, professional-looking map that is wrong by a constant.

Three things have to agree:

• Horizontal datum and realisation. NAD83 is not one thing. The realisation and the epoch matter, and a site that has been surveyed over many years may hold control in more than one. Mixing them shifts features horizontally, which on sloped ground converts into an apparent vertical error.
• Vertical datum and geoid model. This is the one that bites hardest. An orthometric height depends on which geoid model converted the ellipsoidal height, and the older and newer Canadian vertical datums do not agree. The difference is a smooth regional offset, so it does not look like an error. It looks like the contractor built the whole site uniformly low or uniformly high.
• Grid and scale. A site grid with a local scale factor, a UTM zone and a 3TM projection all describe the same ground differently. A surface delivered on a site grid and compared against one on a provincial projection will not line up, and on a long site the mismatch grows with distance from the origin.

The fix is unglamorous: write the datum down. The coordinate system article covers the Alberta options and how to state them so there is nothing to interpret. A surface handed over without a written datum is a surface somebody will have to guess about later.

The design surface is a model, and models have limits

People treat the design surface as truth because it came from the engineer. It is a model, built for a purpose, and it has properties that affect the comparison.

• It stops somewhere. Every design surface has a boundary. Beyond it there is no design, so there is no difference to compute. Difference software will often extrapolate or clamp rather than refuse, and the result is a band of invented numbers around the edge of the map. Ask for the model limits as a polygon and clip the comparison to it.
• It may be the wrong revision. On an active job the design changes. A comparison against a superseded surface is worse than no comparison, because it is actionable and wrong. The revision identifier and its date belong in the deliverable, not in somebody's inbox.
• It may be a different layer than you think. Finished grade, top of subgrade and top of subbase are different surfaces. Comparing an as-built of the subgrade against the finished grade design produces a difference equal to the pavement structure, and somebody will read that as a shortfall.
• Breaklines matter more than point density. A design surface is defined by its breaklines: the crowns, grade breaks and ditch inverts. A triangulation that misses a breakline rounds off the feature, and the difference map will show a stripe along every ridge and swale that is an artefact of the model rather than a condition of the ground.

What the drone actually measured

Photogrammetry reconstructs the top of whatever reflected light back to the camera. That is a more restrictive statement than it sounds, and it is the second big source of surprise on as-built comparisons.

• Water reads as a surface. A ponded area returns the water surface, not the ground under it. Against a design that drains, that is fill where there is actually a hole.
• Vegetation reads as fill. Grass, weeds on a stockpiled area, regrowth on a graded slope. All of it sits above the ground and all of it shows as material that is not there. This is the honest limit set out in what a drone survey cannot tell you.
• Snow reads as fill, unevenly. Worse than vegetation, because it drifts. A capture over snow cover is not a conformance survey, for the reasons in the article on surveying sites under snow.
• Stored material reads as ground. Topsoil windrows, a temporary stockpile parked on the pad, equipment, pipe. Each becomes part of the as-built surface.
• LiDAR changes some of this and not all of it. A LiDAR scan can retrieve ground returns through light vegetation that photogrammetry cannot, which is why it is the right tool on a vegetated site. It does not see through water, and it does not see through a solid stockpile.

The practical response is to time the flight rather than to argue with the physics afterwards. A conformance capture wants bare ground, a dry site and a cleared pad, and it is worth delaying a flight by two days to get them.

Thickness needs two surfaces, and one of them is gone

An elevation check asks where the surface is now. A thickness check asks how much material is between two surfaces, and that needs both of them.

If the pay item is a lift thickness, the underlying surface has to have been captured before it was covered. There is no processing step that recovers it afterwards. This is the single most common irrecoverable gap on drone as-built work, and it happens because the first flight gets scheduled after the first lift is already down.

The fix is to decide at the start which surfaces will be measured and to fly to that list rather than to a calendar. On a staged earthworks job that usually means a capture at original ground, a capture at each subgrade or layer that carries a pay quantity, and a capture at finished grade. The cadence question is covered more generally in how often to re-fly a site.

Tolerance, and whether the survey can adjudicate it

A conformance comparison is only meaningful if the measurement is sharper than the thing being measured against. If the specification allows a given vertical tolerance on finished grade, and the survey's own vertical uncertainty is of a similar size, then a point sitting just outside the band has not been shown to fail. It has been shown to be unresolvable.

That is not an argument against using drone data for conformance. It is an argument for checking the two numbers against each other before anyone relies on the map, and for being explicit about it in the deliverable. The article on what an accuracy figure on a quote actually means covers how to establish the survey number rather than accept it, including the role of independent checkpoints.

Two further points that routinely get missed:

• Relative accuracy is better than absolute. If the question is whether the pad is flat and at the right slope, the comparison depends on internal consistency, which is the stronger of the two figures. If the question is whether the pad is at the right elevation relative to a benchmark, it depends on the absolute figure.
• Accuracy is not uniform across the site. It degrades at the edge of the photogrammetric block and wherever ground control is sparse, which is why ground control placement is a conformance question and not only a cost question.

Reading the difference map without being misled

A cut and fill map is a persuasive-looking object and it is easy to make it say the wrong thing.

• Set the colour ramp to the tolerance, not the data. Software defaults stretch the ramp across the range present in the result. On a site that is almost entirely compliant, that turns normal variation into a map of dramatic reds and blues. Fix the ramp to the specified band and the same data reads correctly at a glance.
• Treat the clipped boundary as the edge of knowledge. Anything outside the design model limits or outside the flown block is not a finding.
• Look for smooth regional trends. A difference that varies gently across the whole site, with no relationship to where machines worked, is the signature of a datum or geoid mismatch rather than a construction outcome.
• Look at the histogram, not only the picture. A distribution centred on zero with a tight spread is a compliant surface. A distribution centred away from zero is an offset, and an offset is usually a frame problem.

The general skill of reading aerial products without over-reading them is covered in how to read an orthomosaic, and much of it transfers directly.

What to supply before the flight

Most of the delay and nearly all of the rework on as-built comparisons comes from information the contractor cannot obtain without the client.

• The design surface as a surface, in an exchange format that carries geometry rather than a picture of it. A drawing set is not a substitute.
• The revision identifier and the revision date of that surface.
• The datum, in writing: horizontal realisation and epoch, vertical datum and geoid model, and the grid or projection including any site scale factor.
• The model limits, as a boundary, so the comparison can be clipped to where a design exists.
• Which layer the design surface represents, and which layer you want the as-built of.
• The specified tolerance, so the deliverable can be built around it and the colour ramp set to it.
• Whether anything stored on the site should be excluded, and whether the site can be cleared and dry on the flight date.
• Existing site control, with the benchmark used and how it was derived, if a previous survey has to be matched.

That list belongs in the scope rather than in an email after the first map. The wider set of clauses worth writing into a procurement is in what to put in a drone survey RFP, and the site-side preparation is in how to prepare your site for a drone survey.

Who can use the result, and for what

A difference map is a management tool the moment it is produced. It becomes a basis for payment only when somebody with the standing to do so certifies the quantity, and that is a professional question rather than a technical one in Alberta. If the comparison is going to support a progress claim, a change order or a final quantity, settle who signs it before the first flight rather than after the number is disputed.

Used for its own purpose, this is one of the highest-value things aerial capture does on an earthworks job: it turns a weekly argument about whether the pad is close into a measured, dated, visual answer. It just needs the design surface to arrive as a surface, on a stated datum, with a revision number on it.