GCP Placement Guide for Survey-Grade Accuracy

The short version

Five to ten ground control points, placed at the corners and perimeter of the flight block, one near the centre, spread across the site's elevation range. Survey more than you need and withhold a third as check points. Aeroyantra achieves 2–3 cm horizontal accuracy on datasets controlled this way, and returns the check-point residuals with the job.

That is the whole method. The rest of this page is why each part matters, because the failure modes are not obvious.

How many

Beyond roughly 15 points the accuracy curve flattens hard. An eleventh point in the middle of a site adds almost nothing; a fifth point at an unconstrained corner can halve the error there. Distribution beats count.

Where to put them

Corners and perimeter first. A bundle adjustment is least constrained at the edges of the block, which is where it bends. An unconstrained corner is the single most common source of large residuals.

One near the centre. This controls doming — the systematic vertical bowl that appears when a block is only constrained at its edges, particularly with a rolling-shutter camera or thin overlap.

Spread them vertically. If every GCP sits at the same elevation, the adjustment has no information about how error varies with height. On a site with relief, put points high and low. This is the part most often skipped and it shows up directly in vertical residuals.

Not in a line. Three collinear points constrain far less than three forming a triangle. Avoid placing everything along an access road because that is where it was convenient to park.

Flat, stable, open ground. Clear of vegetation, away from the base of vertical structures that cast hard shadows or occlude the target at oblique angles. Not on anything that will be moved or driven over between surveying the point and flying.

Making targets that actually work

Size to the imagery. The target must span 8–10 pixels minimum to locate its centre reliably:

Use a high-contrast pattern with an unambiguous centre — a black and white chequered square is the standard because the intersection of the quadrants is geometrically definable rather than eyeballed. Avoid:

• Glossy or laminated surfaces, which blow out to pure white in direct sun and lose the centre
• Spray-painted crosses on dark asphalt, where contrast is too low
• Anything smaller than the table above, on the assumption that you will "zoom in"

Each point should appear in at least five to six overlapping images. A GCP visible in two frames contributes almost nothing to the adjustment.

Control points versus check points

This is the distinction that separates a defensible survey from an assertion.

Control points go into the bundle adjustment. The solution is fitted to them.

Check points are withheld. After processing, you compare their computed position against their surveyed position. That residual is your accuracy.

Software reports an RMS error after adjustment. That figure describes how well the solution fits the points it was given — it is the software grading its own homework. If every marker was used as control, you have no independent measure of anything.

Survey more markers than you need. A two-thirds control, one-third check split is common. Keep the check points spread across the block, not clustered, and include at least one near a corner where error tends to be worst.

A reported RMS of 1.2 cm alongside check-point residuals of 9 cm tells you the adjustment absorbed the error rather than resolving it. You only ever learn that if you held points back.

With RTK or PPK

RTK gives each image a precise camera position, which largely solves the horizontal datum. It does not make ground control pointless.

• Horizontal: the gain from GCPs is small. RTK alone typically lands within a few centimetres.
• Vertical: the gain is real. Vertical residuals commonly improve from 8–12 cm to 3–6 cm with a handful of checked points, because the vertical component of a GNSS solution is inherently weaker than the horizontal.
• Verification: without independent points you cannot demonstrate accuracy to a client or an auditor. For bill certification or compliance work, that is usually the decisive argument.

Practical recommendation: on RTK flights, place three to five points and use them all as check points, feeding none to the adjustment. You get verification at minimal field cost, and the RTK solution stands on its own.

Aeroyantra processes RTK and PPK datasets with no GCPs at all — see processing RTK images without GCPs — and still reports residuals against whatever check points you supply.

Surveying the points

The GCP is only as good as the coordinate attached to it.

• Use a survey-grade GNSS receiver with RTK correction, or a total station tied to a known benchmark.
• Occupy each point long enough to converge. A two-second fix on a rover that has just walked out of tree cover is not a control point.
• Record the coordinate reference system explicitly, including the vertical datum and the geoid model. A file that says "UTM 43N" without stating whether heights are ellipsoidal or orthometric is the most common cause of a survey that is horizontally perfect and vertically wrong by tens of metres.
• Photograph each point in situ with the marker visible. When a residual looks wrong three weeks later, that photo is how you work out whether the target moved.

The CSV format

Aeroyantra accepts ground control as plain CSV — no proprietary format, no conversion step:

name,easting,northing,elevation
GCP01,412355.182,3104221.447,214.882
GCP02,412512.905,3104198.113,213.109
GCP03,412498.331,3104402.776,219.640

State the CRS when you upload. If the file uses latitude and longitude rather than a projected system, that works too, but be explicit about which.

Common failures

All points clustered centrally. The edges of the block go uncontrolled and bend. Very common when the site perimeter is hard to access.

All points at one elevation. Vertical error goes unmodelled. Common on sites where the only accessible ground is the flat area.

Targets too small for the GSD. The centre cannot be located to better than a few pixels, so the "control" is itself uncertain by several centimetres.

No check points. You have a number from the software and no way to know if it means anything.

Undocumented datum. The single most expensive mistake on this list, because it is usually discovered after delivery.

Points surveyed after the flight, on targets that moved. Survey the markers in the same visit as the flight, and photograph them.

Frequently Asked Questions

How many GCPs do I need for a drone survey?

Five to ten ground control points cover most sites up to about 50 hectares: one near each corner of the flight block, one near the centre, and additional points to span the site's elevation range. Add roughly one more point per additional 20-25 hectares. Beyond about 15 points the accuracy gain flattens, and effort is better spent on distribution and on withholding check points than on adding more control.

Where should ground control points be placed?

Place GCPs at the perimeter and corners of the flight block first, because unconstrained edges are where a bundle adjustment bends most, then one near the centre to control doming. Spread them across the site's elevation range rather than all at one level, keep them on flat stable ground clear of vegetation and vertical structures, and ensure each point appears in at least five to six overlapping images.

Do I still need GCPs if my drone has RTK?

RTK removes the need for GCPs to establish the horizontal datum in most cases, but you still want three to five checked points. Two reasons: vertical accuracy benefits measurably from ground truth, and without independent points you have no way to verify the result. Aeroyantra processes RTK datasets with no GCPs at all, and still reports residuals against any check points you supply.

How big should a GCP target be?

Size the target so it spans at least 8-10 pixels in the imagery. At a 2.74 cm GSD that means roughly 25-30 cm across; at a 1 cm GSD, 10 cm is enough. Use a high-contrast pattern with a definable centre, such as a black and white chequered square, and avoid glossy materials that blow out in direct sun.

What accuracy do GCPs actually add?

On a non-RTK dataset, good ground control is the difference between metre-level absolute error and 2-5 cm. On an RTK dataset the horizontal gain is small but the vertical gain is real, typically pulling vertical residuals from 8-12 cm down to 3-6 cm. The larger effect is that GCPs let you prove the accuracy rather than assert it.

Related

• GSD calculator and accuracy guide
• Process RTK drone images without GCPs
• Aeroyantra pricing — $0 to start, $2.50 per credit