Absolute vs Relative Accuracy in Drone Surveying

Why two datasets that are each centimetre-tight can still miss each other by metres — and how to fix it.

Two different promises hiding in one word

“Centimetre accuracy” is sold as one thing but delivered as two. Relative accuracy means points agree with each other: the distance between two corners of a stockpile is right to a centimetre, internal geometry is tight, the map looks flawless. Absolute accuracy means every point is where the Earth says it is — correct in a global reference frame like ITRF2020, overlayable on cadastre, satellite imagery, and last year's survey without shifting anything. A dataset can be brilliantly relative and terribly absolute at the same time, and most accuracy disappointments in drone mapping are exactly that combination.

The distinction is invisible on the day of capture and expensive later. Photogrammetry software reports reprojection errors and tie-point statistics — relative measures. Nothing in the report warns you that the whole block floats 80 cm north-east of reality.

Where the absolute error comes from

It almost always enters through the base station position. RTK and PPK are differential techniques: rovers are positioned relative to the base with centimetre precision, so whatever error the base coordinates carry is stamped onto every rover point, every geotag, every derived surface — perfectly preserved. An averaged single-point base (an Emlid Reach averaging session, or DJI's D-RTK modes at their specified 1.5 m uncalibrated single-point accuracy) can sit a metre from truth; the resulting map is internally centimetre and globally metre. Fly again next month with a slightly different average, and the two flawless datasets miss each other by the difference — the classic symptom that sends teams hunting phantom photogrammetry bugs.

Network RTK bases and known-point setups avoid this by importing coordinates from surveyed infrastructure; PPP-converging bases avoid it by computing absolute coordinates from satellite orbit and clock products directly (mechanics in RTK without CORS).

When relative is honestly enough

Not every job needs the globe. A one-off stockpile volume compared against itself, a construction progress animation, a cut/fill within a single mission — these consume internal geometry only, and an averaged base serves them perfectly well. The test is temporal and spatial isolation: will this dataset ever be compared with another dataset, imported into a design model, or revisited? If genuinely no, relative suffices. The trouble is that “no” has a short shelf life — the client asks for an overlay, the project extends, the machine-control model arrives — and retrofitting absolute accuracy onto relative data means re-surveying or shift-fitting with all its guesswork.

How to get absolute accuracy in practice

Four workable routes, in ascending order of independence. Occupy a known point: perfect where trustworthy monuments exist. Tie to a CORS network: let network RTK position the base (the D-RTK 3 calibration path) — accurate within coverage, subscription attached. Post-process a static log: submit hours of observations to OPUS, AUSPOS, or NRCan CSRS-PPP and receive authoritative coordinates the next day — free, slow, internet required. Self-converge by PPP: the base resolves its own ~1.5 cm ITRF2020 position from L-Band corrections in about three minutes, no infrastructure at all. Modern practice increasingly defaults to the fourth and keeps the third as an audit trail: log RINEX during the mission anyway, and any skeptical reviewer can reproduce your base coordinates independently.

Whichever route, record the metadata that makes coordinates mean something: frame (ITRF2020), epoch, antenna height, and the transformation applied to reach the delivery datum (see ITRF2020 explained).

Checking yourself: the two-checkpoint habit

Verification costs twenty minutes. Before demobilizing, measure two or three independent checkpoints with a rover — points not used anywhere in processing — and compare against the final surface or orthomosaic. Residuals inside a few centimetres prove the entire chain: base coordinates, corrections, geotags, photogrammetry, export. Residuals of a consistent half-metre in one direction diagnose a base-position offset instantly, while scattered large residuals point at FLOAT flying or processing trouble (triage steps in the FIX checklist). Two checkpoints will not satisfy a boundary-survey statute, but they transform accuracy from an assumption into a measurement — which is the entire point of the absolute/relative distinction.

The vocabulary sellers use — decoded

Spec sheets exploit the ambiguity fluently. “RTK accuracy 1 cm + 1 ppm” is a relative figure — rover relative to base, silent about where the base is. “Survey-grade accuracy” usually means the same. “Absolute accuracy” claims should name a frame and a method: absolute in what, established how? DJI is refreshingly explicit — publishing 1.5 m for uncalibrated single-point D-RTK operation and 1 cm + 1 ppm only after network RTK calibration — and that honesty is a model for reading everyone else: whenever a centimetre claim appears, ask what anchors it. If the answer is “your known point” or “your CORS subscription”, the centimetre is conditional. If the answer is “the receiver converges itself by PPP to ITRF2020”, the centimetre is intrinsic.

The same decoding applies to photogrammetry reports: RMS reprojection error, tie-point sigma, and even GCP residuals measure internal consistency. Only independent checkpoints measured against an absolute reference measure truth.

Case file: the stockpile that shrank

A quarry flew monthly volumetrics for a year with an averaged base position re-established each visit. Individual surveys were superb; month-over-month volume deltas swung inexplicably by thousands of cubic metres. The cause: each month's average landed decimetres from the last, tilting and shifting the comparison surface — volume differences amplified the misalignment across 20 hectares. The fix took one change: a self-converging PPP base whose coordinates were identical (to centimetres) every visit. Deltas stabilized to loader-ticket agreement immediately. The lesson generalizes to any longitudinal work — construction progress, erosion monitoring, rehabilitation compliance: absolute accuracy is what makes time-series subtraction legitimate (more mining patterns in stockpile surveys).

A field checklist that keeps you honest

Before capture: write down what anchors the base (PPP convergence, known point, network tie, or averaged — and if averaged, accept the consequences deliberately). During capture: note frame, epoch, antenna height. After capture: shoot two independent checkpoints; compare; file residuals with the deliverable. Three moments, five minutes total, and every accuracy conversation with a client moves from adjectives to numbers.

One-line takeaway

Relative accuracy makes a map agree with itself; absolute accuracy makes it agree with the world — and the base station's own coordinates are where the difference is decided, before the drone ever takes off.

Further reading

The frame side of this story — epochs, plate motion, and why ITRF2020 makes absolute coordinates portable across years and borders — continues in ITRF2020 explained; the hardware side, how a base establishes absolute coordinates with no infrastructure at all, is covered in RTK without CORS.

Frequently asked questions

My map looks perfect but sits 1 m off the cadastre. What happened?

Classic relative-only symptom: the base position was averaged rather than absolute, and its offset transferred to every point. Re-establish the base absolutely (PPP, known point, or OPUS/AUSPOS) and the shift disappears.

Do GCPs give absolute accuracy?

Only as absolute as the instrument that measured them. GCPs shot from a rover tied to an offset base inherit the same offset. Anchor the base first; then a couple of checkpoints suffice.

Is repeatability the same as absolute accuracy?

Effectively yes over time: if every visit is positioned in the same global frame, datasets align by construction. Averaged bases produce a new frame each visit — precise, unrepeatable.

Which frame should I deliver in?

Capture in ITRF2020, deliver in whatever the client's GIS expects via a documented transformation. Never adjust raw data by eye to force an overlay.