Ground Control, Checkpoints, and Accuracy

Center for Geospatial Analytics at North Carolina State University

Corey White

Objectives

• Explain the relationship between flight GSD and the horizontal and vertical accuracy a UAS survey can actually achieve
• Design ground control and checkpoint schemes, and defend the difference between the two
• Choose between GCP-based and RTK/PPK georeferencing, and know what each still needs
• Read an Agisoft and a WebODM processing report and state the accuracy of the products with the right numbers
• Decide, with arithmetic, whether a given flight can answer a given question: counting trees or measuring millimeters of erosion

This is the accuracy lecture the SfM topic has been pointing at since Topic 2: 2A said checkpoints are the honest number, 2B said quote the checkpoint RMSE, and today is where those numbers come from and what they permit. Everything is anchored to the reports students produced themselves in Assignments 2B and 2C and to the checkpoint data collected with the Emlid rover at the Lake Wheeler flight. The lab runs the full validation on the flight’s own products.

The report number is not the accuracy

A worked example straight from the ASPRS standard (American Society for Photogrammetry and Remote Sensing 2024):

• Your DSM fits the checkpoints with RMSE_V1 = 1 cm. The report says so.
• The checkpoints themselves were surveyed with an RTK rover: RMSE_V2 = 2 cm.
• The accuracy you may claim is the quadrature sum:

\[\text{RMSE}_V = \sqrt{1^2 + 2^2} = 2.24 \text{ cm}\]

Edition 2 rule: checkpoint survey error can no longer be ignored; reported accuracy includes it. Your product is never more accurate than the ruler you measured it with

Open with the arithmetic because it reframes the whole hour: the software prints a fit, not an accuracy. Edition 1 dodged this by requiring checkpoints three times better than the product, so their error was negligible; Edition 2 relaxed that to two times and made you carry the term instead. For our lab the rover repeatability from the field day is RMSE V2. This is also the first appearance of the lecture’s habit: every accuracy claim must say what it was measured against and how good the ruler was.

What GSD buys you

Section one: the floor that pixel size sets, and why you rarely stand on it.

GSD sets a floor, not the accuracy

	Best case (strong control, good geometry)	Flat terrain envelope	Complex terrain envelope
Horizontal RMSE	about 1 x GSD (asymptote 0.66 x)	1 to 3 x GSD	1 to 7 x GSD
Vertical RMSE	1.5 to 2 x GSD	1 to 4.5 x GSD	1.5 to 5 x GSD

• Envelope columns from Jiménez-Jiménez et al. (2021); best-case column from Sanz-Ablanedo et al. (2018)
• Vertical error runs about 2.5 x the horizontal error, consistently (Sanz-Ablanedo et al. 2018)
• Vertical never reaches 1 x GSD, no matter how much control you add
• With uncorrected navigation-grade georeferencing, errors reach meters, tens of GSD or more: the pixel size stops mattering entirely

The envelope is Jimenez-Jimenez et al. 2021; the asymptotes are Sanz-Ablanedo et al. 2018, from 3465 GCP combinations over a 1200 ha block. Teach the multipliers as a band, not a constant: the gap between the floor (1x) and what you get is governed by georeferencing and systematic error, not pixel size. Ask the class: your Lake Wheeler flight had roughly 2 to 3 cm GSD; what vertical accuracy should they predict before the lab? Write the predictions down; the lab checks them.

Planning with GSD

• ASPRS recommends source imagery GSD of 0.5 to 1.0 x the target horizontal accuracy class: a 5 cm class wants 2.5 to 5 cm GSD (American Society for Photogrammetry and Remote Sensing 2024)
• Read backwards: expect RMSE_H of 1 to 2 x GSD from a well-run project
• The standard is explicit that this is a planning heuristic, not an accuracy statement

Workflow: start from what you must measure, derive the accuracy that requires, derive the GSD, and only then the flying height (the Topic 3A formula)

This connects 3A’s GSD formula to a requirement instead of a preference. The design direction runs measurement to accuracy to GSD to altitude, never the other way. The standard’s own caveat is worth quoting: the table should not be used to reference product accuracy. It tells you what to fly, not what you achieved.

Ground control

Section two: how many points, where, and the difference between the points you fit and the points you judge with.

How many GCPs, and where

• Count in flight units: about 3 to 4 GCPs per 100 photos; below 1 per 100 accuracy degrades fast (Sanz-Ablanedo et al. 2018)
• Horizontal stops improving past about 3 per 100; vertical keeps improving to about 4 per 100
• Distribution beats count: an even triangular spread is twice as accurate as a poor layout with the same number
• Edge coverage first, then an even interior; more points past that buy better error structure, not lower RMSE (Dai et al. 2023)

The per-100-photos rule travels better than per-hectare rules because it scales with overlap and altitude automatically. Dai et al. 2023 is the modern synthesis: after the first few GCPs, RMSE barely moves but the spatial autocorrelation of the residual error keeps improving, which matters for differencing. Our 9/16 field design (3 to 5 targets on a 78-photo class of flight) sits right on the recommended density.

Control is not validation

GCPs (control): used in the bundle adjustment; the model is fitted to them

Checkpoints: surveyed the same way, withheld from processing entirely; the model is judged against them

• Error reported on control points overstates accuracy by 3 x (dense control) to 4 to 8 x (sparse control) (Sanz-Ablanedo et al. 2018)
• The adjustment was told to fit those points; their residuals are a statement about the fit, not the map
• ASPRS: checkpoints must be independent, excluded from the aerial triangulation, and away from any GCP (American Society for Photogrammetry and Remote Sensing 2024)

The 3 to 8 times overestimation is the single most important number in this section, and it is why the field protocol withheld most of the surveyed points. If students remember one distinction from today, it is this one; the midterm will ask it.

RTK, PPK, and what they still need

Section three: direct georeferencing moved the field in the last five years; here is the current state and the honest recipe.

Direct georeferencing today

Evidence	Result
ASPRS Addendum V (American Society for Photogrammetry and Remote Sensing 2024)	PPK reaches 1 to 2 cm horizontal, 2 to 4 cm vertical “when used properly”; RTK typically 5 to 10 cm
RTK aircraft, zero GCPs, purpose-built flight (Stott et al. 2020)	Vertical RMSE 6.6 cm over 2 km of river; adding 5 GCPs did not improve it
PPK versus GCP workflows (Zhang et al. 2019)	PPK matched GCP-workflow accuracy (MAE about 2 cm, RMSE about 3 cm); some missions carried a vertical bias that one GCP removed, improving vertical accuracy 20 to 30 percent
Repeat surveys (Nota et al. 2022)	Co-alignment of epochs gives sub-2 cm relative accuracy; ground control in at least one epoch still anchors absolute Z

Stott’s caveat is the teaching point: the zero-GCP result required a double grid flown in opposing directions at 20 degrees off nadir. Against the sparse target checkpoints, vertical errors still clustered spatially up to 0.14 m even though the dense validation found no consistent systematic error, so distributed checkpoints stayed mandatory. Zhang’s one-GCP result is the canonical citation for the vertical bias fix: the bias is near constant, so one known point kills it. Nota matters for the change-detection topic later: co-align the epochs and put control in one of them.

Standard’s caution

Defaults: RTK or PPK for georeferencing, 1 or 2 GCPs to control the vertical datum, and every other surveyed point withheld as a checkpoint

• The ASPRS standard is more conservative: GCPs remain “highly recommended” and should take priority over airborne positions in the adjustment, because camera centers have fewer epochs and weaker geoid handling than survey-grade ground occupations (American Society for Photogrammetry and Remote Sensing 2024)
• Both are right: the journal results are conditional on strong flight geometry; the standard writes for contracts and courts

Present the tension honestly instead of resolving it: recent peer-reviewed results justify the light-control recipe when the flight design is strong; the standard assumes you must defend the number to a client. A student consultant should know both positions and say which regime they are operating in. Our lab uses the recipe: 3 targets as control, everything else checkpoints.

The receiver in your hands

• Base within 0.5 mm: an Emlid Reach RS2 base gave DEM vertical RMSE of 1.94 cm versus 1.89 cm for a Trimble Alloy with a choke-ring antenna (Famiglietti et al. 2021); the price gap between those two bases is roughly an order of magnitude
• Antenna beats receiver: a low-cost dual-frequency board matched geodetic instruments in network RTK only when paired with a geodetic-grade antenna; with a patch antenna it needed open sky (Janos and Kuras 2021)
• The base coordinate governs absolute accuracy: the rover measures centimeters relative to the base, and every point inherits the base position (the field protocol’s Methods A, B, and C)

This slide connects the literature to the hardware students actually held. The Famiglietti comparison is the honest Emlid-class citation; the receivers have since improved. The publishable-gap card is real: nobody has published an RS4-versus-state-RTN evaluation, and the class dataset is a credible short note. Recruit interest for it here.

Systematic error

Section four: the error that does not average away, and that differencing cannot forgive.

Doming

• Nadir-only image blocks with self-calibration cannot separate radial lens distortion from surface shape
• Camera orientations off parallel by only 2 degrees produce a dome of about 0.2 m over 100 m (James and Robson 2014)
• Fixes, in order of power: a known camera model with self-calibration off; oblique imagery (practical plans cut the deformation by one to two orders of magnitude; a convergent orbit reached under 2 mm in simulation); two flying heights; well-distributed control
• The best camera angles in the tested 0 to 35 degree range were 20 to 35 degrees, worth nearly 50 percent better accuracy in high relief (Nesbit and Hugenholtz 2019)
• Caution: a gently tilted camera (under 15 degrees) can make things worse by exciting a different distortion correlation, especially on consumer wide-angle cameras with onboard lens correction (James et al. 2020)

Recall the pitfalls table from 2A; here is the mechanism and the numbers. The 2020 caveat is the one practitioners miss: the popular “just tilt the camera a bit” advice backfires below 15 degrees on exactly the cameras students fly. More overlap, incidentally, does not fix doming; James et al. 2020 showed relative insensitivity to overlap, a common myth.

Why systematic error is the enemy

• Random error averages away in aggregates; a 2 cm systematic bias does not
• In modern change detection, propagated uncertainty is dominated by low-magnitude, spatially correlated or systematic error, not by point precision (Anderson 2019)
• A noisy point cloud is a smaller problem than a quiet dome

Blunder rule (American Society for Photogrammetry and Remote Sensing 2024): any checkpoint discrepancy beyond 3 x the target RMSE is a blunder to investigate, not average in; mean error should stay under 25 percent of the target RMSE

The mean-error rule is the practical doming detector in a report: a mean checkpoint error far from zero, or a mean above a quarter of the RMSE, says the surface is biased, not noisy. Students apply exactly this test to their DSM residuals in the lab.

Reading the reports

Section five: the week’s stated purpose. Bring up the students’ own 2B and 2C reports side by side on screen for this section; the slides give the map, the reports give the territory.

The same facts, two dialects

Quantity	Agisoft Metashape report	WebODM quality report
GSD	Survey Data: ground resolution	Overview: average GSD
Coverage and photos	Survey Data: coverage area, aligned images	Overview: area, reconstructed images
Overlap	Camera locations and image overlap figure	Survey Data overlap heatmap
Camera model	Calibrated values with uncertainties and residual plot	Camera parameters, per-band
Camera position error	Camera Locations table: X, Y, Z error	GPS/geolocation details, 3D errors
Control fit	Ground Control Points table: control RMSE	GCP errors section (when GCPs used)
Checkpoint error	Check Points rows in the same table	Absent unless checkpoints were declared

Walk each row on the projector with a real pair of reports. Two traps to name: both tools print camera-position errors that students mistake for map accuracy, and WebODM only reports what it was given, so a project processed without declared checkpoints reports no independent accuracy at all, which does not mean the map is accurate. The lab’s comparison table is this slide filled in with their numbers.

What neither report tells you

• Whether the checkpoints were independent (the software trusts your labeling)
• The survey accuracy of the control and checkpoints, so neither can do the quadrature for you
• Whether the error is spatially structured: a mean near zero can hide a dome that cancels
• Anything about the parts of the site with no checkpoints at all

Report vocabulary (American Society for Photogrammetry and Remote Sensing 2024): report RMSE_H and RMSE_V (RMSE_3D if asked); the old 95 percent confidence reporting is gone from Edition 2; 30 checkpoints is the standard’s minimum for a statistical claim, and Section 7.16.1 gives the honest wording when, as in a class lab, you have fewer

Section 7.16.1 matters for every project the students will run this decade: small sites cannot afford 30 checkpoints, and the standard’s answer is truth in reporting rather than silence: report the RMSE, state the checkpoint count, and do not claim the statistical confidence the sample cannot support. The lab report requires exactly that sentence.

What can you measure

Section six: the payoff. The same flight either answers the question or it does not, and the arithmetic decides.

The level of detection

For change between two surveys of similar quality, at 95 percent confidence:

\[\text{LoD} = 1.96 \sqrt{\sigma_{z1}^2 + \sigma_{z2}^2} \approx 2.8\,\sigma_z\]

Worked example for a good flight: 3 cm GSD, strong control, so \(\sigma_z \approx\) 1.5 to 2 x GSD \(\approx\) 5 cm:

\[\text{LoD} \approx 2.8 \times 5 \text{ cm} \approx 14 \text{ cm}\]

• A 30 cm gully clears the threshold; 2 cm of sheet erosion is invisible
• Thresholding is for gross change only; it biases net change, where random errors already cancel (Anderson 2019)
• Spatially variable LoD can exclude from a quarter to more than half of observed volumetric change (Wheaton et al. 2010)

The LoD arithmetic is the lecture’s destination: every accuracy number from the previous fifty minutes exists to be plugged in here. Wheaton’s Feshie numbers (24 to 31 percent of change excluded in dry years, 57 to 59 in wet) show the price of thresholding; Anderson’s correction says when to pay it: gross change yes, net change no, and error propagation is not thresholding. Zhang et al. 2019 reached sub-5 cm detection thresholds with PPK flights below 45 m, which is what chasing micro-topography actually requires.

Trees or erosion

Question	Binding constraint	Verdict at 3 cm GSD, 5 cm sigma-z
Count the trees	Detection in the orthomosaic	Easy; accuracy barely matters
Measure canopy height	Canopy and ground surfaces, not GSD: flying height 80 to 120 m showed no significant effect on tree height RMSE (1.8 to 3.2 m) (Grybas and Congalton 2022)	Expect meters of error regardless of pixel size
Detect a 30 cm rill or gully	LoD about 14 cm	Yes, cleanly
Quantify 2 cm sheet erosion	LoD, dominated by systematic error	No; needs sub-2 cm sigma-z, near-ground flights, and bias control

The asymmetry is the thesis of the whole topic: for trees the limiting error lives in the canopy and the terrain model underneath it, so buying more GSD buys nothing; for micro-topography the limiting error is exactly the accuracy chain this lecture built, and every link (base coordinate, control, calibration, systematic bias) shows up in the final LoD. Students argue these verdicts with their own numbers in the lab’s part three.

Key terms

Checkpoint: surveyed, withheld, judges the map; also called check point, validation point, independent test point

Direct georeferencing: camera positions from RTK or PPK doing the work of control; also called GNSS-aided or GCP-free

Level of detection (LoD): the smallest elevation change distinguishable from error at a stated confidence

Doming: broad systematic DEM deformation from the self-calibration ambiguity; also called bowling when inverted

RMSE_H, RMSE_V, RMSE_3D: Edition 2’s reporting quantities, replacing RMSEx/RMSEy/RMSEz and the 95 percent statistics

Quadrature: accuracies add as the square root of summed squares; the product carries the checkpoint survey error

The synonym lists matter because students will search the literature and hit every variant. This is midterm vocabulary.

Wrap-up

Assignment 3B: Validate the Lake Wheeler flight: establish the base coordinate three ways, evaluate the flight’s orthomosaic and DSM against the checkpoints you collected, and rule on what the flight can measure

Midterm 11/4: the control versus checkpoint distinction, the quadrature rule, and the LoD arithmetic are all fair game

Recap the chain: GSD sets the floor, georeferencing and systematic error set the distance above it, checkpoints measure that distance honestly, the quadrature keeps the claim honest, and the LoD converts the whole chain into a yes or no about the question you actually care about. The lab data is their own field work from 9/16, which is the point.

References

American Society for Photogrammetry and Remote Sensing. 2024. ASPRS Positional Accuracy Standards for Digital Geospatial Data, Edition 2, Version 2. American Society for Photogrammetry; Remote Sensing. https://publicdocuments.asprs.org/PositionalAccuracyStd-Ed2-V2.

Anderson, Scott W. 2019. “Uncertainty in Quantitative Analyses of Topographic Change: Error Propagation and the Role of Thresholding.” Earth Surface Processes and Landforms 44 (5): 1015–33. https://doi.org/10.1002/esp.4551.

Dai, Wen, Ruibo Qiu, Bo Wang, et al. 2023. “Enhancing UAV-SfM Photogrammetry for Terrain Modeling from the Perspective of Spatial Structure of Errors.” Remote Sensing 15 (17): 4305. https://doi.org/10.3390/rs15174305.

Famiglietti, Nicola Angelo, Gianpaolo Cecere, Carmine Grasso, Antonino Memmolo, and Annamaria Vicari. 2021. “A Test on the Potential of a Low Cost Unmanned Aerial Vehicle RTK/PPK Solution for Precision Positioning.” Sensors 21 (11): 3882. https://doi.org/10.3390/s21113882.

Grybas, Heather, and Russell G. Congalton. 2022. “Evaluating the Impacts of Flying Height and Forward Overlap on Tree Height Estimates Using Unmanned Aerial Systems.” Forests 13 (9): 1462. https://doi.org/10.3390/f13091462.

James, Mike R., Gilles Antoniazza, Stuart Robson, and Stuart N. Lane. 2020. “Mitigating Systematic Error in Topographic Models for Geomorphic Change Detection: Accuracy, Precision and Considerations Beyond Off-Nadir Imagery.” Earth Surface Processes and Landforms 45 (10): 2251–71. https://doi.org/10.1002/esp.4878.

James, Mike R., and Stuart Robson. 2014. “Mitigating Systematic Error in Topographic Models Derived from UAV and Ground-Based Image Networks.” Earth Surface Processes and Landforms 39 (10): 1413–20. https://doi.org/10.1002/esp.3609.

Janos, Daniel, and Przemysław Kuras. 2021. “Evaluation of Low-Cost GNSS Receiver Under Demanding Conditions in RTK Network Mode.” Sensors 21 (16): 5552. https://doi.org/10.3390/s21165552.

Jiménez-Jiménez, Sergio Iván, Waldo Ojeda-Bustamante, Mariana de Jesús Marcial-Pablo, and Juan Enciso. 2021. “Digital Terrain Models Generated with Low-Cost UAV Photogrammetry: Methodology and Accuracy.” ISPRS International Journal of Geo-Information 10 (5): 285. https://doi.org/10.3390/ijgi10050285.

Nesbit, Paul Ryan, and Christopher H. Hugenholtz. 2019. “Enhancing UAV-SfM 3D Model Accuracy in High-Relief Landscapes by Incorporating Oblique Images.” Remote Sensing 11 (3): 239. https://doi.org/10.3390/rs11030239.

Nota, E. W., W. Nijland, and T. de Haas. 2022. “Improving UAV-SfM Time-Series Accuracy by Co-Alignment and Contributions of Ground Control or RTK Positioning.” International Journal of Applied Earth Observation and Geoinformation 109: 102772. https://doi.org/10.1016/j.jag.2022.102772.

Sanz-Ablanedo, Enoc, Jim H. Chandler, José Ramón Rodríguez-Pérez, and Celestino Ordóñez. 2018. “Accuracy of Unmanned Aerial Vehicle (UAV) and SfM Photogrammetry Survey as a Function of the Number and Location of Ground Control Points Used.” Remote Sensing 10 (10): 1606. https://doi.org/10.3390/rs10101606.

Stott, Eilidh, Richard D. Williams, and Trevor B. Hoey. 2020. “Ground Control Point Distribution for Accurate Kilometre-Scale Topographic Mapping Using an RTK-GNSS Unmanned Aerial Vehicle and SfM Photogrammetry.” Drones 4 (3): 55. https://doi.org/10.3390/drones4030055.

Wheaton, Joseph M., James Brasington, Stephen E. Darby, and David A. Sear. 2010. “Accounting for Uncertainty in DEMs from Repeat Topographic Surveys: Improved Sediment Budgets.” Earth Surface Processes and Landforms 35 (2): 136–56. https://doi.org/10.1002/esp.1886.

Zhang, He, Emilien Aldana-Jague, François Clapuyt, Florian Wilken, Veerle Vanacker, and Kristof Van Oost. 2019. “Evaluating the Potential of Post-Processing Kinematic (PPK) Georeferencing for UAV-Based Structure-from-Motion (SfM) Photogrammetry and Surface Change Detection.” Earth Surface Dynamics 7 (3): 807–27. https://doi.org/10.5194/esurf-7-807-2019.