The project specification calls for imagery with a 5cm GSD. The drone is flown at the right altitude. The processing software states that the GSD is approximately 5 cm, and the orthomosaic appears extremely sharp.
Does that indicate that the survey is accurate to 5cm?
Not at all, and it is in that gap between the two figures that many disputes, redos, and unpleasant calls from clients arise.
What GSD Actually Means
The ground sample distance is the actual size on the ground of a single pixel, it being calculated from the flight altitude, the resolution of the sensor, and the focal length. If the GSD is 5cm, then each pixel in the image corresponds to a 5cm by 5cm area of the ground. That's all; the GSD informs you about how well the image can resolve details but gives no information as to where those details are actually located on the Earth.
Why GSD Is a Resolution Measurement, Not an Accuracy Measurement
Imagine that GSD refers to the sharpness of a photograph and accuracy to whether the photograph is correctly marked with the correct position; you could have a very sharp image of the wrong area or a sharp image of the right area that is still offset by a metre from where the labels say it should be. GSD limits the level of detail that you can in theory resolve — your actual accuracy can never be finer than your GSD — yet achieving that limit involves a completely different set of decisions, none of which are related to GSD.
GSD vs Horizontal Accuracy
The horizontal accuracy refers to the degree to which the position of a feature on a map corresponds to its actual position on the ground, as measured in the X/Y plane. According to published research on UAV photogrammetry, the typical horizontal root mean square error (RMSE) is about 1 to 3 times the GSD when the terrain is flat and as high as 1 to 7 times the GSD in the case of complex topography. To put this another way, a survey with a 5cm GSD can have a horizontal error ranging from 5 to 15 cm on level ground and much worse on hilly or uneven terrain — even though there is nothing wrong with the imagery itself.
GSD vs Vertical Accuracy
Vertical accuracy is less tight than horizontal accuracy, a finding that is confirmed by the same research: on flat ground the typical vertical RMSE is between 1 and 4.5 times the GSD, and on difficult terrain it is between 1.5 and 5 times the GSD. A separate study on the number of GCPs and their distribution found that horizontal and vertical RMSE values of about 3x GSD and 5x GSD respectively could be achieved with at least three well-distributed GCPs. In the case of a 5cm GSD survey, this means that the vertical error could realistically range from 5cm to 25cm depending on the type of terrain and on how the ground control was managed. That is precisely the reason why the elevation values in the DSM and DTM should be examined more closely than the orthomosaic's visual sharpness implies.
RTK/PPK vs GCPs
RTK and PPK provide the drone with an exact and corrected position at the instant each photo is taken, greatly reducing the difference between the ground sample distance and the actual accuracy achieved. However, while RTK/PPK corrects the position of the camera, it does not necessarily correct the position of objects on the ground. GCPs still serve as the direct and independent link between the model and the real-world coordinates obtained by surveying. The two methods are not substitutes for one another but rather different levels within the same accuracy issue: RTK/PPK improves the geometric accuracy, and the GCPs confirm and fix it to ground truth.
Why Checkpoints Matter
A checkpoint is a point that has been surveyed and is deliberately excluded from the processing stage — it is not used when constructing the model but is only used later to test it. The importance of this difference is greater than it may at first appear: as one can see from a drone survey forum discussion on this exact topic, an RMSE calculated without independent checkpoints is in fact only measuring the consistency of the camera position with respect to GPS, something that has nothing to do with ground accuracy. An RMSE based on GCPs that have also been used to build the model shows how well the model fits the data you provided it — it does not indicate how accurate the model actually is. Only by using an independent checkpoint can the real question be answered. We went through this exact distinction between checkpoints and GCPs as part of a real project pipeline in yesterday's breakdown of how drone survey processing actually works.
How GCP Distribution Affects Accuracy
The vertical error has been found to rise as the distance from the nearest GCP increases, indicating a measurable pattern rather than a random one. If a number of GCPs are closely grouped in one corner of a site, it is possible to achieve good accuracy values right near that group while having considerably lower accuracy in all other areas — figures that may still average out to an outcome that appears acceptable in an overall report, even though they conceal the actual local errors. We have also looked at the practical aspect of this issue in how many GCPs a particular site actually requires.
RMSE and What It Actually Tells You
Root Mean Square Error gives a single figure that shows the usual size of the error over a group of compared points — but since it is an average, it does not guarantee accuracy. It is possible for a dataset to have a good overall RMSE at the same time as containing certain areas with considerably poorer accuracy, especially in regions far from the GCPs or in terrain where the model could not reconstruct cleanly. While RMSE is a useful summary figure, it does not prove that the accuracy is the same throughout the entire site.
Why a Visually Perfect Orthomosaic Can Still Be Geometrically Wrong
The pipeline's different sections are responsible for achieving good visual quality and geometric accuracy. A sharp, well-blended and seamless orthomosaic is a sign of good overlap, good lighting, and accurate feature matching — yet none of these factors ensure that the underlying geometry is correctly positioned in real-world coordinates. The orthomosaic can appear perfect and yet still be shifted, warped, or inconsistently scaled over the area if the georeferencing and camera calibration beneath it are not reliable.
How Terrain Affects Accuracy
The accuracy obtained is considerably different according to the terrain, even when the GSD, the number of GCPs, and the flight parameters are the same. The research mentioned above indicates that the horizontal RMSE multipliers are about double when going from flat to complex terrain, and that vertical error increases with both changes in elevation and with distance from the GCPs. A flight plan and GCP strategy that has been validated on level ground cannot guarantee the same level of accuracy when applied to a site that has real variations in elevation.
Camera Calibration
An uncalibrated or badly calibrated camera model leads to systematic errors throughout the entire dataset — this consistent bias does not appear as clear distortion in any individual image but instead subtly affects the overall accuracy of the model, no matter how small the GSD is. This is one of the reasons why the camera optimization stage of the processing procedure is just as important as the flight itself.
Ground Sampling Distance vs Relative/Absolute Accuracy
Accuracy can be divided into two separate concepts which are easily confused. Relative accuracy refers to internal consistency — whether or not two points on the model are correctly situated in relation to each other. Absolute accuracy means how correctly the entire model matches the true, independently surveyed ground coordinates. It is possible for a model to have good relative accuracy — being internally self-consistent, clean, and well-aligned — yet still be absolutely incorrect, that is, shifted as a whole from its true ground position. GSD has no direct connection with either of these; it is a resolution limit that applies under both types of accuracy.
How to Prove a Survey Meets Its Accuracy Specification
Meeting the GSD specification and meeting the accuracy specification are two different claims, and only one of these can be verified by referring to the imagery. To prove accuracy, it is necessary to use independent checkpoints that have been kept separate from the processing, calculate the RMSE specifically from those checkpoints rather than from the GCPs employed to create the model, and, if possible, carry out a distribution check to ensure that the errors do not increase substantially away from the GCP locations. This kind of documentation is what transforms an accuracy claim into one that a client can check for themselves rather than having to just trust it — exactly the same method that settled an actual dispute at a live dam site, as described in how we actually proved a GSD specification was met after a client raised an objection to it.
The Actual Takeaway
The GSD indicates the level of detail that a survey could in theory achieve, but it gives no information as to the actual position of that detail on the Earth's surface. Two surveys can have the same GSD and yet differ by a factor of ten in their real positional accuracy, the amount depending completely on the number of GCPs, the way the GCPs are distributed, the checkpoint validation, the camera calibration, and the terrain — none of which are reflected in the GSD value. Therefore, if your project specification merely states a required GSD, it's important to find out what accuracy standard is actually being guaranteed, since these are two separate figures that answer two different questions. If you need both the resolution and the verified positional accuracy that a project actually requires, then this is exactly the kind of planning that topographic survey and drone imaging and processing services are designed to carry out from the beginning.
