Ground Control vs No Control

Our office has often wondered how accurate aerial surveys done by RPAS are. There are many reports been done by surveyors comparing datasets of orthomosaics and DEM to conventional survey methods. We have done so numerous times over both large and small datasets for validation to both ourselves and to other users. Nearly all reports have shown that there is very good correlation between conventional ground methods and RPAS photogrammetric surveys.
However, there seems to be a lack of information as to how accurate data derived from photogrammetric processes compares when excluding ground control and using only GPS location and IMU data from an autopilot. Many photogrammetry programs allow for flight logs that contain GPS tracks, IMU logs and camera trigger events to be included as a reference to determine the image centers giving a model a spatial dimension.
We recently undertook a comparison drone survey over a site to see the difference between final end products of a survey controlled only by a flight log and a survey corrected by addition of control points from ground survey.

The site chosen for the comparison is located in Kwazulu-Natal, South Africa. This active construction site is approximately 2km by 1km. Control was initially placed with Leica GPS1200 survey grade GPS. The control markers consisted of paint marks on tar or hardened surfaces and arranged plastic bags weighted with soil. A total of 40 control points were distributed evenly over the site and measured to acceptable limits (0.02m error) via RTK methods. The survey was referenced to our local approximation of WGS84.
After the control was placed and coordinated, a fixed wing RPAS of our own design was deployed with a 24mp RGB camera onboard, triggered to capture images at set intervals as determined by the flightplan. The flightplan was derived in the open source Mission Planner using the criteria of 65% sidelap, 70% forward lap, 170m AGL.
The flight took 35min to complete with 576 images recorded. A flight log from the Pixhawk autopilot running Arduplane 3.4 firmware was extracted in .LOG format.

Once the requisite data was collected, the raw images and flightlog were process using Pix4D ver.2.0.100 on a Windows 10 I7 machine. 2 separate projects were processed, one with only the flightlog used to reference the images, the second using the flightlog to initially reference the images but later adding control points to further refine the model.
No-Control: 568 out of 576 images were used in the calibration. The remaining images were rejected as they appeared to be shots taken during banking by the aircraft. After processing, the absolute error found from the interpolated camera positions to computed was X:5.58m, Y:3.80 and Z:3.70. This is within the expected range given the size of the site, the delay in recording the image and instrument inaccuracies.
Control: The same results as above were computed, but ground control were added to the model and then realigned. The final RMS error for all the control was 0.02m. Othophotos and DEM were generated for both projects and exported for analysis

Over 80 points where selected on each corresponding orthophoto evenly over the site. These were chosen at random so as to not be influenced or weighted by the existing control points. The points were coordinated in 2d and compared.
The overall shift from control to no-control did not appear visually to be shifted in a planar fashion with a rotation but that a scaling had also been introduced. The 2d mean shift between points averaged at dX:0.19m, dY:0.04m, however the range varied as much as rX:10.97m, rY:10.94m. The distribution of the shift is exaggerated at the edges of the model. A Helmert transformation of the data shows that a scaling factor of 0.995 and a rotation of 0.0619 degrees would need to be applied to gain similar dimensions to the control data.

Comparison points
Comparison points

This information is surprisingly good as it shows that despite a rotation the linear measurements over the entire 2km of the site are only 1% exaggerated, the exaggeration being weighted towards the edges of the model.

Horizontal errors
Horizontal errors

The same points were compared with their respective DEM data and Z values derived. From this information, values where expected to have an even shift over the site as the control datum for each survey system would have varied. The average shift was found to be 3.01m however the range varied 13.82m. This means for this data set, the error range is 9%. The data spread appears to be exaggerated in the center and edges of the model. Cross section were drawn at random to confirm this inconsistent shift.

Vertical errors
Vertical errors

The reasoning as to why there is not a simple shift in the X,Y and Z is due to a number of factors. The main being that the GPS and IMU sensors on-board the autopilot are not sensitive and accurate enough. An uncorrected GPS will never give results below a meter in accuracy, while the IMU needs to be perfectly aligned with the camera. Then there is the time delay between when an event is recorded and when the image is finally recorded after the delay of signal, light metering and focus lag. For an aircraft travelling at 15m/s, a shutter lag of only 500ms means that the camera will have already travelled 7.5m before it is recorded. Then there are the internal optics of camera lenses to consider that are not perfectly aligned and can distort the final model.
The implications of not using control for othometric surveys is sever. Having only the data sheet from the photogrammetric software saying that, locally, the model is accurate, may be far from the truth on the ground (where it matters). This data has to be confirmed and quantified by other methods as sensitivity and accuracy of on-board sensors is not sufficient.
As previously mentioned, the scaling of a model as 1% over 1000m is 10m and is acceptable for certain types of survey. When combined with the vertical element, the variance is not so predictable. For calculating slopes, volumes and gradients, having such a large data range (9%) is far too much for any accurate estimates.

Inconsistent cross sections

Volumetric work for instance requires accurate and consistent heighting. To accurately determine a volume, a base layer needs to be determined. After earthworks, the area is resurvey using the same datum and the models are subtracted from eachother to determine a volume. Having no control causes errors and uncertainties as the datums will be different for each survey and that we have identified uncertainty within the model.
This is not to say that the data is useless. There may be instances where this data is acceptable such as the beginning stages of land use planning or basic modelling. Often it does not make financial sense to include ground control for simply an orthophoto of a site. In these cases, the data is perfect as it is, but it cannot be transformed into a more accurate rendering of the terrain
It is expected that in the future, more robust and accurate sensors will be available to surveyors. These are already making their way to market with higher frequency GPS chips that can be post processed to centimetre level, accurately timed IMU sensors and event triggering timers for cameras. It will be a question of time before placing control before an aerial survey is done, becomes redundant.
In conclusion, it is irresponsible of a surveyor to not include control with their survey when undertaking quantifiable work. The uncertainty of the data issued can cause the client to make costly mistakes when doing earthworks or other quantity calculations. It is not sufficient to issue information as accurate and absolute because the processing software says it is. The surveyor needs to remember his training and apply first principles when undertaking this type of survey as there is certainly a potential for exaggerated results.