Drone Based GCPs Part 2
Part 2 of our investigation into using small drones as mobile ground control points for aerial surveying.
Please see Part 1 for an introduction.
Here we tackle bullet point 2:
Can the drone's on-board GPS be used to record the GCP locations? To what accuracy?
So we set up a data logger to record GPS readings from the most common drone GPS, the 3DR module that uses the UBlox NEO-7n chipset.
Relevant features include:
GPS and GLONASS
5 hz update rate
LNA and SAW filter
25x25x4mm ceramic patch
TT2963, (a similar set-up to the main image above) and recorded readings once per second for 12 hours. NOAAs online tool was used to convert between NAD 83 (2011) and WGS 84 (1674) to compare the GPS reading with the survey monument control location.We placed the GPS unit and data logger together on a survey monument,
This first plot shows raw GPS Longitude readings, the survey control location, and a running median of the GPS Longitude readings.
Typical "Pro-sumer" GPS units with WAAS are expected to read a typical accuracy of around 1m. The WAAS system guarantees no worse than 7.6m 95% of the time. We were also curious if we could increase this accuracy (especially the 95% confidence interval) by taking a median over a long period of time. The plot above seems to confirm this hypothesis, where the the median approaches the survey location line quickly over the first hours and then stays a short distance away while the real time longitude readings continue to fluctuate about the median.
This motivates the question how long do you have to take readings to achieve a desired confidence interval? So we crunched the data and plotted the 2D horizontal RMS error over a number of rolling intervals throughout the entire 12 hour period.
These results are significant. The raw instantaneous RMS fluctuates rapidly with the ionosphere and other typical sources of GPS error. However, as the rolling median window gets longer these mostly symmetrical fluctuations are damped more and more until the improvement tapers off at around 2 hours. The table below contains the final results for each rolling median window length and the 95% confidence RMS.
So for this 12 hour period the GPS performed about as expected with a little better than a meter accuracy on average and 2.22 meters 95% RMS. What's interesting is the relationship between "soak time" and the RMS value.
In this particular 12 hour window we had an uncorrected 2D RMS of 2.22 meters. If we were to park the drone for 2 hours and take the median of all the GPS readings during that time we would have had a lat/long better than 0.66 meters accuracy 95% of the time. If you can wait 2 hours it looks like you can bring your accuracy up by a factor of four by using these statistical methods. So perhaps an expensive DFGPS base station is not required if only slightly sub meter accuracy is required.
Of course more testing is required to confirm this particular result on this particular day is representative. But it is definitely exciting that a typical drone GPS's accuracy can be increased simply by letting it sit a for a few hours and taking an average or median of the logged positions.