How accurate are drone mapping surveys?

Drones are being used more and more to provide maps to support construction, development, mining, agricultural and other commercial activities. Queensland Drones has been providing aerial mapping services for over two years now, so it seems like a good time to discuss their sources, accuracy and reliability.

Aerial drone maps are being used to support all kinds of planning, design and assessment

First, let’s be clear about what drone maps are not … they are not a substitute for a site survey. Drone mapping can be amazingly accurate, but before you rely on drone-based maps to plan earthworks or building foundations or others structural and engineering uses where accuracy is critical, they need to be ground-truthed or otherwise assessed by a qualified surveyor.

When we create aerial survey maps, we put a lot of effort into ensuring maximum accuracy by using high quality drones with good GPS and IMU technology to capture the imagery, but also using a lot of ground control points (GCPs) to level and align the photogrammetric data in post-processing. Some drone operators don’t use GCPs at all, but rely on a 3rd party processing platform like DroneDeploy to provide the accuracy. These platforms provide pretty outputs, but not accuracy.

What does “accuracy” mean for drone surveys?

The basis for photogrammetry is knowing where each photo is positioned in space – its latitude, longitude and altitude. Let’s start with the concept of how a drone positions itself and obtains this data in the first place.

Drones typically rely on commercial satellite technology to know where they are and to navigate between locations. Technologies like GPS and GLONASS, which use satellites positioned 20-40,000 km up in space. While satellites can provide fairly accurate positioning, communication between the satellites and the drone can be affected by all kinds of things including solar radiation (sunspot flares), heavy cloud, tall buildings, forests (where the satellites are low on the horizon) and more. Despite all this, the GPS/IMU in a modern drone can figure out where the drone is in space down to a metre or two.

A metre or two … sounds pretty accurate until you want to use it to place a post or a foundation or a bridge support, in fact almost anything that really matters needs to be placed where it needs to be placed, not a metre or two away.

Drone surveys convert overlapping aerial photos into a 3d point cloud

When we talk about drones and accuracy, we are really talking about two different things and both of them are quite important.

First, there’s relative accuracy (sometimes called the “local” accuracy).  This refers to how accurate one point is on a map compared to another point. It’s typically used to measure distance and volume, so you might want to know how long a section of road is, or how much material is left in a stockpile, or how much will need to be removed to dig a hole this big.

You can make a lot of local or relative measurements without the map necessarily being accurate about its position in space, because the measurements are relative to points within the generated map itself. So as long as the map is not distorted in some way, relative accuracy can be useful.

The larger the mapped area, the less the error in relative accuracy is critical to the outcome. For example, if we assume that relative accuracy is one metre for a given aerial mapping technology and we’re measuring 25km of fence line, an error of 1m or so is not terribly important. But if we’re measuring a retaining wall 10m long, an error of 1m is pretty significant.

The second measure is absolute accuracy (sometimes known as global accuracy). This refers to how well a given point on a photogrammetric map corresponds to the actual position of that point if it were measured, for example, by a surveyor.

If we use the road example above, if the drone imagery says the starting point of the road is Latitude 152.34567, Longitude -267.4321 and is 120m above mean sea level, and a surveyor’s measurement says something different, then we have an absolute accuracy error. In our surveys we strive to capture at least two or three previously surveyed markers with high accuracy to use as benchmarks for the rest of our ground control points. This ensure high levels of absolute accuracy.

Absolute accuracy is the more critical measure of how useful a drone mapping survey will be when used for design, construction or establishing property boundaries. It’s less critical when surveying for agriculture, for example.

How accurate are drone mapping surveys?

Well, that depends on how they are planned and executed, and whether you’re concerned with local or global accuracy. It helps if you understand how drone mapping surveys are created in the first place.

A drone is flown over a defined area, usually in a grid pattern, capturing a photo either every couple of seconds or every couple of metres depending on the settings. Each photo should overlap the neighbouring photos so that photogrammetry software can find common points and stitch them into a single (orthophoto) image. This single image is the basis for everything that comes after it, so it needs to be pretty damn good.

How good that imagery is depends very much on how the operator plans and executes the flight. Time of day, weather conditions, light consistency, height above ground, image overlap, wind, nearby sources of radio interference, sunspot activity and more can and will affect how well the imagery can be aligned and stitched.

Two orthographic maps overlaid on a Google Earth map show a large error in global and local accuracy

If you go down to Harvey Norman and buy a Phantom 4 drone, then plan and execute a mapping flight with the DroneDeploy app, the relative positional data of each image you capture could be less than a metre horizontally and 1.5m vertically, or it could be a lot worse.

The image above shows two photogrammetric maps (one near infrared, one RGB colour) captured at the same time, in the same place, but which when overlaid on Google Earth are actually nowhere near where they should be in space and show curvature which indicates that even their local accuracy is unreliable.  The problem with these maps is that a strong solar storm was warping the earth’s magnetic field when these were captured, interfering with the drone’s ability to position itself using GPS satellite data.

The local accuracy of the maps above are probably still within a couple of metres, although the global accuracy of the map in relation to the position of these objects on the earth is about 10-20 metres.

At this point I need to introduce another measure of accuracy, known as Ground Sampling Distance (GSD). This is a measure of local accuracy which comes from the height at which the drone is being flown and the characteristics of the camera in the drone.

A method used by Pix4D to calculate ground sampling distance (courtesy of Pix4D)

GSD is a measure of the size of an object that would be clearly visible in a single image captured by the drone. For example, at a GSD of 5cm, an object 5cm x 5cm in size (let’s say an iPad) on the ground would be represented by a single pixel in the photo. You would know the object is there (assuming its colour is significantly contrasting with its surroundings), but you could not tell what it is. At a GSD of 2.5cm, the iPad would be represented by four pixels in the photo and you may be able to see what shape it is, but you would still struggle to know what it is. If you were able to achieve a GSD of 1cm, the iPad would be represented by 25 pixels in the photo and there’s a good chance you could see what it is.

A DJI Phantom 4 Pro drone being flown at 120m above the ground will achieve a maximum GSD of 3-4cm per pixel. At this height, a single photo will represent a ground area of about 210m x 140m at 20MP (megapixels). By comparison, a DJI Phantom 3 Pro at the same altitude has a GSD of 5-6cm.

To get a better GSD (and therefore a higher resolution mapping outcome), the drone needs to be flown at a lower altitude. If the Phantom 4 Pro above is flown at 60m instead of 120m, the GSD goes down to 1.9cm per pixel. Problem is, each image now covers just 100m x 70m, so you get a lot more of them for the same mapped area. The more images there are, the harder they are to stitch and the more local features are needed to ensure accurate stitching.

As a rule of thumb, the best local horizontal accuracy you can achieve with a drone will be equivalent to the GSD, so let’s say 3-4cm for a Phantom 4 Pro. Vertical accuracy is another thing entirely and is generally about 3 times GSD, so for the Phantom 4 Pro at 120m the vertical accuracy is about 9-12cm minimum. Remember, this is local accuracy so it only relates to measurements within the orthophoto itself, not to positional accuracy on the ground. What this means is if you are calculating the volume of a stockpile of material, the edges of the stockpile as seen in the stitched photo may be out by 3-4cm in any direction, and the height of the stockpile may be out by 9-12cm. This is assuming everything else is perfect.

This is a good point to introduce another concept, that of overlap. This is the amount by which each photo overlaps each of its neighbours. When the photogrammetry software tries to stitch the photos into a single image, it relies on finding identifiable common points in many photos.

How photos overlap each other in a typical photogrammetry mapping mission (courtesy of DroneDeploy)

In the photo above you can see how each photo overlaps its neighbours to both the front and the side. This is called, logically enough, frontlap and sidelap. If there is a shed in one of these photos, chances are the same shed will be visible, in part or in full, in many of the neighbouring photos. The software uses this common information to join (stitch) the photos together.

Generally speaking, the lower you fly the more overlap you need to have in your photos, because common objects appear in less photos. Again, generally speaking, the more uniform the surfaces in the imagery, the more overlap you need. So if the surface is a suburban neighbourhood with lots of homes, lower overlaps are OK because there are lots of unique features to pick out. But if the surface is, say, bare earth or forest, you need much higher overlaps to find common points.

How well the images overlap and how many common points the software can find between them will directly affect the quality of the final stitched image (orthophoto) and therefore its usefulness in terms of local accuracy measurement.

What about global accuracy?

You may recall I mentioned earlier that global accuracy is about whether the X, Y and Z coordinates (latitude, longitude and altitude) of any given point in an orthophoto really correspond to that point’s position on earth if measured by a surveyor.

To achieve global accuracy in drone mapping you need a lot more than a drone and a flight planning app. You need to be able to align key points in the imagery to their actual locations on the ground using precision positioning equipment.

There are essentially two types of positioning equipment used for drone mapping surveys – PPK (which is about more accurately positioning the drone in space) and RTK (which is more about knowing the accurate position of objects in the photos). There is a raging debate about the comparative accuracy of each system, but in truth they can both be important to improve the absolute (global) accuracy of drone maps.

There are a lot of factors that can affect the global accuracy of drone mapping data, including the type of camera used (resolution, edge distortion, etc), how stable it is held in the air (does it move around when photos are being captured), the GSD and overlaps (which were discussed earlier), the weather (cloud, light, variability of light, solar flares, etc), the terrain (tall buildings, forests, mountains, etc) and the GPS signal reliability.

But the most important factor is being able to relate the objects in each photo to their position on the ground with a very high level of accuracy. For this we use Ground Control Points.

Why are these so important? If you think about software trying to stitch multiple photos together based on the position of common objects in those photos, the software is sometimes going to distort a photo to make it fit with the other photos. Every time it does that, positional errors are being introduced for other nearby objects. Unless the software can know the absolute position of these objects on the ground, it cannot correct these distortion errors.

GCPs are essentially points on the ground which have been marked in a way that they are clearly visible in the photos captured by the drone, and for which the position (latitude, longitude and altitude) have been accurately measured using a meaningful measurement system. Both RTK and PPK can be used to measure the position of points on the ground. RTK is, in my opinion, better because it can use a network of precisely located base stations to correct the measured position of a point in real time (hence the name, Real Time Kinematic). PPK (Post-Processed Kinematic) can be more precise, but requires additional processing after each flight.

Basic ground control points marked using white crosses on the ground

The number and distribution of the ground control points in a drone mapping survey will significantly contribute to the global (absolute) accuracy of the mapping that results. Aligning the ground points to known survey points will also improve their reliability and their acceptance for survey purposes.

There are many ways that small accuracy errors can be introduced into a drone mapping survey. For example, if the centre of a GCP is a line 2cm wide (so it can be easily identified in the photos), then positioning that point in the imagery will have a potential error of 1-2 cm just because of the thickness of the line. Add this to the positional accuracy of PPK or RTK (typically 1-2cm at best) and you have a total potential error of 4-6 cm already.

Drone maps created using good quality drone equipment, careful flight planning, commercial grade RTK ground control points and commercial grade processing software can potentially be accurate to around 2-3cm horizontally and around 5-6cm vertically.  Because we prefer to err on the side of caution, we claim our aerial mapping accuracy as <10cm horizontal and <15 cm vertical, which is generally sufficient for most purposes. Anyone claiming accuracy higher than this probably does not understand what they are doing.

Note always that drone operators are not surveyors. While our mapping can be created at very high levels of accuracy, it should be certified by your surveyor before being used for engineering purposes.

If you would like to enquire about drone mapping for your projects, please contact us.