There are two ways to think about accuracy:
This is the accuracy of a point on the map relative to other points within the map/model. When asking questions like: "How long is this fence?", "What's the area of this field?" or "What is the volume of this stockpile?", it doesn't really matter where the map/model is in the world, as long as the map/model is consistent with itself. This is the important measure for length, area, and volume.
For example, if we created a giant map of Antartica and wanted to measure the shortest distance from the South Pole to the Arctic Ocean we would need high relative accuracy in our map, meaning the size and shape of Antartica are correct so we find and measure the shortest distance. Say that distance is really 300 miles and we measured it as 300 miles and 20 feet, the relative accuracy of the shortest distance between the South Pole and the ocean would be within 20 feet.
Absolute accuracy, on the other hand, is the degree to which the calculated position of a point on a map corresponds to its actual position in a fixed coordinate system in the real world. If a map has a high level of global accuracy, the latitude and longitude of a point on that map will correspond fairly accurately with actual GPS coordinates.
This is important when you need a high degree of confidence that the lat/long and elevation measurements of each point on the map are correct when comparing with the real world, e.g. when comparing to geo-referenced design documents for a construction project or conducting property boundary surveys.
In our giant map of Antartica example, the absolute accuracy of the location of the South Pole is the difference between where it is on our map and where the real South Pole. If the map shows the South Pole 2 feet to the left of where it should be, then the absolute accuracy of that point is within 2 feet.
Note: your results may vary
These are examples to give you a rough idea, please don't quote these as expectations to support, there are many factors (see below).
There are a number of factors which affect accuracy, however, as a rough rule of thumb, giving a typical drone, modern GPS (able to receive signals from several constellations), and atmospheric conditions this is what you may expect. These numbers come from comparisons to DJI Phantom 3 data, flown with DroneDeploy at 250ft compared to Ground Control Points.
The horizontal accuracy within a map largely depends on the Ground Sampling Distance (GSD, i.e. number of pixels per centimeter) of your data. You can expect the local error to be around 1 to 3 times the average GSD of the data.
As mentioned below, this largely depends on where you are on Earth, and what GPS receiver you have. Using a standard GPS on a DJI Phantom 3, you can typically expect to have around 1 meter (3 ft) horizontal accuracy. So if you draw a circle around you with a 1m radius, and give someone your GPS location, they're expected to be somewhere within this circle.
The rule of thumb is that Absolute Vertical Accuracy will be around 3 times worse than the horizontal so we would expect around 3 meters.
You can radically improve your Absolute (or Global) GPS accuracy by using Ground Control Points (GCPs) and Checkpoints or Differential GPS systems (RTK, PPK, etc.). More on this below, but these can increase your Absolute accuracy to maximum of around 2-5cm horizontally, and 4-8cm vertically.
There are primarily 3 ways you can process a map with DroneDeploy:
With no GCP's
With GCP's and checkpoints
In each of those situations, the RMSE accuracy values are describing different things. We’ll be using the glossary in the Accuracy Report to define each of these situations, which you can find in the 'i' information section of each map.
When we process without GCP's, the RMSE number is the is the 'Optimized Camera Location XYZ RMSE' What does that mean?
The camera location XYZ root mean squared error (RMSE) is the average image location error in the XYZ axis for all images in the map. The image location error is the difference between the image location that is recorded by your drone's GPS and the corrected image location that is calculated during map processing. Therefore, as an example, a 10ft Camera Location XYZ RMSE means that on average in the XYZ dimension image GPS locations were 10ft away from the corrected image locations.
*Please note that camera location error does not correspond to the true accuracy of a map. For example, poor GPS conditions can cause large camera location errors but if images are properly collected the processed map will still be highly accurate. To truly measure map accuracy you must include checkpoints or an object with known dimensions which can be measured in the processed map to check for differences.
To keep it short, it’s how 'off' the drones GPS tags are compared to where we needed to place the images to complete the model. This data point we can't extrapolate too much from. Anecdotally, it is a useful number to serve as a basic sanity check of the accuracy map. But we can't quantifiably infer accuracy from it.
What about when we use GCP's? What are GCP's? GCP's tie the digital and the physical by adding known, fixed points to the map area. When we use GCP's the RMSE value is the ‘GCP XYZ RMSE’.
The ground control point (GCP) XYZ root mean squared error (RMSE) is the average GCP location error in the XYZ axis across all the processed GCPs. The GCP location error is the difference between the GCP location as measured by your precision GPS device and the corrected GCP location that is calculated during map processing.
*Please note that GCP location error does not correspond to the true accuracy of a map. This is because the corrected GCP locations are calculated using a mathematical estimation which is weighted so corrected locations will be close to the measured location. To truly measure map accuracy you must include checkpoints or an object with known dimensions which can be measured in the processed map to check for differences.
It's pretty much the same thing as last time, except swap out images with GCP points. Again, we can't extrapolate the true accuracy of the map from the GCP error because we've forced the GCP's to go to a particular spot.
These GCP points are weighted - we know where they need to go. They aren’t 'control variables'. We need to remove that weight in order to truly test accuracy. So, checkpoints are what you can use to actually validate the accuracy of your map to a survey grade level.
When we use checkpoints, the RMSE value is the ‘Checkpoint XYZ root mean squared error’.
The checkpoint XYZ root mean squared error (RMSE) is the average checkpoint location error in the XYZ axis. The checkpoint location error is the difference between the checkpoint location as measured by your precision GPS device and the correction checkpoint location that is calculated during map processing.
*Please note that checkpoint location error is a measure of the absolute accuracy of your map at that single location. Systematic errors can cause a map to have large checkpoint location errors and a low absolute accuracy but the map may still have a high relative accuracy. For example, shifting all the points in a map by 1ft in the Z direction will create a 1ft checkpoint location error without effecting the relative map accuracy.
Checkpoints are unweighted GCP's. As in, they aren't used in processing whatsoever. Because of that, they're a control variable.
For example, we know that checkpoint 1 needs to go to coordinate 'A, B, with elevation C'. But, we aren't forcing it to go to that spot. We are letting the 'natural processing forces' float the checkpoint to wherever it happens to go.
And because we let that natural process occur, the difference between where that checkpoint is in real life and where we say it is on our map is the actual absolute accuracy for that point
If you have a number of evenly distributed checkpoints throughout the map, you can use their accuracy values arrive at an average absolute accuracy of the entire map. It’s never perfectly describing the accuracy of every single point in the entire map, but will paint an accurate picture.
Drones can create highly accurate data, but the accuracy is dependent on a number of factors
- Camera: Better and bigger sensors have less noise, less blur, and less of a rolling shutter effect, which will produce better data
- Lens: Less lens distortion (barreling or fisheye) will produce better data
- Drone: Drones with gimbals keeping the camera pointing correctly will produce better data
- Altitude: The higher you fly, the less accurate things like elevation will be as it's harder to tell the relative difference between two distances the further you are away from it
- Image resolution: Higher resolution imagery will produce better data because there's more information to match against
- Number of photos: The more images, the more GPS locations we have to work with. This produces less error because of the Law of Large Numbers
- Higher Overlap in imagery: The higher the overlap in images, the more key-points we can detect, and the more GPS data we'll have for each pixel, increasing accuracy
- Atmospheric Conditions: GPS is affected by: Atmospheric Conditions (temperature, air density, pollution, clouds), Ionospheric conditions, Solar Flares
- Buildings: Tall structures block GPS signals, as well as reflect them (commonly called the "Urban Canyon") causing multi-path interference which causes inaccurate data
- Location on Earth: There are several GPS constellations (see below), and where you are on Earth limits the number of satellites you can
- GPS in the US
- GLONASS in Russia
- GALILEO in Europe
- BeiDou-2tf in China
- NAVIC/IRNSS in India
- Where GPS and GLONASS are the only Global systems, the others only have local visibility
- GPS Receiver: Different GPS receivers are able to listen to different constellations (listed above). Being able to accept more signals give more sats to use for positioning which improves accuracy.
- Differential GPS: RTK, PPK etc. has access to corrections of the GPS data which radically improves accuracy (meters >> cm).
Note: This isn't the ACTUAL Absolute Accuracy, but a proxy to this.
Below, we show you where to view the accuracy on DroneDeploy. This isn't the actual accuracy, but a proxy to the Absolute Accuracy of your data.
Without Ground Control Points, it's impossible to know the precise Absolute Accuracy, but we work out an expectation on the accuracy by comparing the difference of where the drone thought it was (using it's GPS), and where it needed to be in order to make the images overlap and stitch appropriately.
Choose a map from your dashboard, then click map information icon to view the map's information:
The map information panel brings up details about the map. Here, if you scroll down, you'll see the total area, the native resolution, and the accuracy of the map:
The RMSE is the Root Mean Squared Error, this gives you the mean accuracy across all the X, Y, Z dimensions.
Reminder: Accuracy shows the average error between where the GPS said the camera was, and where DroneDeploy calculated the camera needed to be in order to make the overlapping images stitch. Where the expected camera locations are off by a large number, we expect lower data accuracy.
Each of the items listed above in What Affects My Accuracy? section shows an area to improve.
The biggest improvements can be made by:
Using a Differential GPS
Differential GPS systems like RTK and PPK will radically increase your accuracy. We currently support the SenseFly eBee imagery with embedded RTK on Business and Enterprise plans. This typically produces an absolute horizontal accuracy of within 1 to 3cm. Read more about our compatibility with eBee RTK here.
Using Ground Control Points
Ground Control Points processing with DroneDeploy add another layer of location data to the map, rather than relying solely on the GPS of the drone. With added GCPs and Checkpoints, we expect 1-5 centimeters of accuracy. This is greatly dependent on your Ground Sampling Distance, i.e. the number of pixels/cm.
Adding Ground Control Points
DroneDeploy can incorporate Ground Control Points and Checkpoint processing for Business or Enterprise customers. Please visit our Processing GCPs with DroneDeploy guide for a full overview of the GCP process.
RTK (Real Time Kinematics) and PPK (Post Processed Kinematics) are procedures that allow for greater accuracy when it comes to GPS Surveying. Kinematic is a common term used in traditional GPS surveying methods where the receivers are in constant motion. To process RTK data, you will need both an RTK Base Station and a RTK Rover, in this case, the Phantom 4 RTK. RTK processing does not require post-processing of the data to obtain an accurate position. This allows for in-field surveying and allows the surveyor to check the quality of the measurements without having to post-process the data.
PPK surveys are similar to RTK surveys, but the images are not corrected in real time. It usually involves placing a stationary base station over a known control point, or a monument to allow for geolocation. GPS data is then simultaneously collected by the base station and the drone as it flies. That data is then downloaded from the base station, and the images are post-processed using a GPS software. These images can then be uploaded to the DroneDeploy Map Engine for processing.
When choosing between RTK or PPK methods, the pilot or surveyor needs to make a choice between productivity and accuracy of the resulting imagery. The RTK workflow can be a very quick way to obtain accurate imagery, but relies on a real-time connection to produce accurate maps, whereas a PPK solution takes more time to set up, but can rely on it’s signal backup data to ensure the flight performs as expected.