The SCSI has funded research into the accuracy of drone-collected survey data. PETER KINGHAN reports.

Drones, unmanned aerial vehicles (UAVs), remotely piloted aircraft systems (RPAS), unmanned aircraft systems (UASs), or simply unmanned aircraft, are one of the most exciting technologies to surface over the last few years. Slowly, drones are starting to be used in a wide range of industries as the commercial world opens its eyes to their potential. The drones are one thing, but it’s the sensors that can be mounted on them that provide the greater potential.
For geomatics surveyors it’s not the drones, or the sensors, but the structure from motion photogrammetric software (SfM) that has proven a game changer for the profession. This SfM software, largely automated, has made it significantly easier, and quicker, to produce high-quality digital surface models (DSMs) and ortho-image mosaics from imagery acquired with variable orientations and overlap, and from cameras that have not been rigorously calibrated. The visualisation of the final SfM product is really impressive, but is the inherent data quality as impressive? Just because the final product looks right, if not carried out properly it may not necessarily be right and if the ‘old age’ lessons of photogrammetry have not been practised, then the results definitely won’t be right.

Potential sources of error
Surveyors talk and work in millimetres. One way to ensure silence in a room of geomatics surveyors is to mention that you had a 100- millimetre error in your traverse. This size of error raises eyebrows! So, considering this sensitivity, what are the potential sources of error from drone surveys, how accurate is the survey information obtained, and what are the ‘tricks of the trade’, or the important things to take into consideration when producing survey data from drone-collected imagery?

Just because the final product looks right, if not carried out properly it may not necessarily be right and if the ‘old age’ lessons of photogrammetry have not been practised, then the results definitely won’t be right.

The SCSI has funded research to find the answers to these questions, and to ensure that members are up to speed and fully in the picture on this relatively new technology and, more importantly, on SfM software. Some potential sources of error from RPAS- and SfM-produced products include:

■ image quality,scale and geometry;
■ number of images;
■ number and location of ground control points (GCPs);
■ GCP configuration;
■ GCP ground type, e.g., soft/hard;
■ GPS accuracy;
■ human error(as with all survey methods);
■ flight height/pixel resolution;
■ light conditions;
■ availability of textures;
■ overlap;
■ type of terrain;
■ inaccurate GPS/inertial measurement unit (IMU) sensors;
■ shutter speed;
■ camera lens alignment;
■ accuracy of the real-time kinetic(RTK)GPS used to survey the GCPs;
■ the real exposure end time of an image;
■ the GPS timestamp tagged to it; and,
■ wind speed.


FIGURE 1: A fixed-wing survey grade drone was used to carry out the SCSI research.

Considering this non-exhaustive list of potential errors, is millimetre accuracy survey data possible from photography collected by drones? In short, yes, but it’s not simply a case of throwing the drone in the air, downloading and processing the data. The SCSI research, which was carried out on a quarry in Westmeath using a fixed-wing survey grade drone (Figure 1), produced different datasets, with the same imagery, using a number of different ground control points and configurations. The results of the research indicate that accuracies of 1-2m root-mean- square-error (RMSE) are achievable when no GCPs are used. Accuracies of 0.08m RMSE were achieved when 24 GCPs were used spread evenly across the area. All data was collected at 120m flight height, with greater accuracy achievable from data flown at a lower flight height (although high spatial resolution does not necessarily imply correspondingly high spatial accuracies) (Figure 2).


FIGURE 2: Digital surface model (DSM) from the survey site, a quarry in Co. Westmeath.

Lessons
There are a number of learning outcomes from this research that will help to improve the accuracy of drone-collected survey data:

■ flight height/ground sample distance (GSD) affects accuracies (lower flight height = higher pixel resolution = (potentially) more accurate results);
■ GCP ground conditions (preferably located on hard surfaces);
■ number of GCPs (the more the better for higher accuracy);
■ preferred GCP configuration (one GCP every three baselines along
the flight direction and every two baselines perpendicular to the
flight direction);
■ human error (can be an issue, as with all survey methods);
■ number of images(increasing the number of images produces
denser meshes and improves model accuracy); and,
■ vertical imagery is geometrically weak(strengthen image geometry by obtaining oblique imagery in addition to the vertical dataset acquired for object coverage).

The key for practitioners is to be open and clear when it comes to achievable accuracies and for users of drone-collected survey data to be clear on what survey specifications they require as different workflows will dictate the accuracy achievable.
Safe flying!
There is an SCSI UAV working group, which has produced a consumers’ guide advising on the regulatory requirements, etc., of drone operators. The guide is available at: https://www.scsi.ie/documents/get_lob?id=1020&field=file.

Peter Kinghan

Chartered Minerals Surveyor, Chartered Geomatics Surveyor and Chair of the Minerals Professional Group