I recently did some testing of the accuracy of the export data of photogrammetry when using images tagged with RTK metadata compared to those constrained by GCPs and the combination of the two. I also looked at how the density of GCPs affected accuracy.
I found RTK only missions were capable of achieving the same accuracy as those fully constrained by GCPs.
And for my 100 acre test site, 10 GCPs were required to reduce vertical accuracy down to expected levels. However only 4 were required for horizontal accuracy convergence.?ÿ
?ÿ
Please feel free to share your thoughts
?ÿ
You would expect that significantly more points would be required to hold vertical accuracy compared to horizontal accuracy. The advantage of using GCPs is that you KNOW their absolute value is correct, but with RTK control you are relying on the residual values to be small. The comparison also depends on how you define your accuracy.
Probably the biggest problem is that the software always tends to "blend" the heights, so it is difficult to get it to produce a sharp gradient change (such as along a low retaining wall). This can result in levels for a metre of two either side of the step being affected. It's therefore a good idea to keep GCP positions in an area which is a uniform grade, otherwise this can reflect into the accuracies.
Practically you shouldn't use the aerial information for high accuracy levels such as road kerb lines; I resolve the problem by running ground survey along them as part of a general check on the derived levels. Drop those strung lines into the ground model and remove the original aerial points in the vicinity (say within 1 metre).
You would expect that significantly more points would be required to hold vertical accuracy compared to horizontal accuracy. The advantage of using GCPs is that you KNOW their absolute value is correct, but with RTK control you are relying on the residual values to be small. The comparison also depends on how you define your accuracy.
Probably the biggest problem is that the software always tends to "blend" the heights, so it is difficult to get it to produce a sharp gradient change (such as along a low retaining wall). This can result in levels for a metre of two either side of the step being affected. It's therefore a good idea to keep GCP positions in an area which is a uniform grade, otherwise this can reflect into the accuracies.
Practically you shouldn't use the aerial information for high accuracy levels such as road kerb lines; I resolve the problem by running ground survey along them as part of a general check on the derived levels. Drop those strung lines into the ground model and remove the original aerial points in the vicinity (say within 1 metre).
Yea, it took about twice as many GCPs to resolve horizontal accuracy. I actually didn't expect the horizontal accuracy to resolve as fast as it did, this was quite surprising to me.
"The advantage of using GCPs is that you KNOW their absolute value is correct, but with RTK control you are relying on the residual values to be small."
I guess you can take multiple shots on the same GCP to "KNOW' it is correct but with the RTK tagged imagery the redundancy is far more dense. Instead on 10 control points you can have images 800 (in this example) with accurate absolute locations with the RTK location "accuracy" tagged in the metadata of the image. Of course it goes without saying that in either case you must have check shots to verify the accuracy of the export data.
I guess one could argue a theory behind "needing half as many GCPs with RTK" is that you can now use those as check points and not actually hold them as control at all. But then I don't know if I would call them GCPs as they should not be held as control if you were to use them as checks.
I have noticed the rounding of horizontal edges, typically what I've seen was a few cm up to a decimeter and since it's not practical to set GCPs along the top and bottom of every vertical face (I haven't witnessed that fixing this issue) that doesn't really affect the comparison here as far as I can see. It is what it is when it comes to drone photogrammetry, that being said there are techniques to reduce this affect. Both with how the images are taken (oblique imagery) and how they are processed (higher quality settings & less depth filtering).
Thanks for the reply
I have noticed the rounding of horizontal edges, typically what I've seen was a few cm up to a decimeter and since it's not practical to set GCPs along the top and bottom of every vertical face
Yes, that's typical of the sort of error I have found. You don't have to set GCP along top and bottom of faces - just keep them a couple or more metres away, so that the ground model smoothing by the algorithms doesn't affect the adjustment onto the GCP.
I am not licensed. And perhaps I am totally missing the mark.
I only have about 5 years experience with GNSS RTK but 4 of them required tying into a known benchmark. Not a previously GPSed point but a USGS geodetic benchmark. I have noticed that you generally have your most precise shot within a few seconds of occupying a point. The longer you sit the farther off it goes.
Longer occupations are not for increasing precision but for increasing confidence. (Which holds up in court)
The more observations you take at a point. The wider the the range of measurements you will receive. Then you can take the standard deviation and build a confidence level. at 5 minutes you are I think like 3 sigmas in, or 97% confidence that the true value lies in that range. Where as with only 10 observations you are only at 67% confidence level (most people just say 70%).
Why don't we take the initial observation since it usually is so much more precise? Because there is a greater chance it is wrong.
So the GCPs are CYA points. Efficiency wise you could do without, and usually be right.
I would go so far to say... do it pure RTK as an initial flight, then double check it trigonometrically when you are doing your ground based data collection, road stripes, Top of walls etc. However I am not a licensed surveyor and what I think does not really matter compared to what the law defines. (Like why not use an inverted rod to measure up when differential levelling? Because someone in the office might make a mistake when you put a minus sign infront of a FS measurement? A double negative means add. That's Junior High math at worst...)
EDIT: There was an article posted her before about how to use oblique ortho IIRC to eliminate tall grass and shrubs. (Mind blown)
Bless you for your research. Keep everyone honest 🙂
Why don't we take the initial observation since it usually is so much more precise? Because there is a greater chance it is wrong.
one initial observation can't be precise. it's only one observation. could be 24 satellites, multiple data points, one observation. or very well could be 10 satellites, lots of multipath, high PDOP, ETC ETC.
BUT not precise.
I have noticed the rounding of horizontal edges, typically what I've seen was a few cm up to a decimeter and since it's not practical to set GCPs along the top and bottom of every vertical face
Yes, that's typical of the sort of error I have found. You don't have to set GCP along top and bottom of faces - just keep them a couple or more metres away, so that the ground model smoothing by the algorithms doesn't affect the adjustment onto the GCP.
This is amplified with bridges that appear bent or rounded due to the large elevation difference in the relationship of the mostly flat bridge deck and the terrain adjacent to the abutments and below.
when setting my Ortho targets for our large format fixed wing systems distance from the vertical relief to objects like buildings etc was minimum of 25 meters or more, unless you're collecting it in the canyons of NYC, that's a whole other level of overflight planning for covering the project to get accurate data to process and cleanup.
I have noticed that you generally have your most precise shot within a few seconds of occupying a point. The longer you sit the farther off it goes.
Do you have data to back up the claim that relative accuracy and positional outliers do not improve after "a few seconds"? Because there are quite a few research papers that demonstrate the opposite.
That's like saying we shouldn't observe F1/F2 observations with a total station, because after the F1 observation if we change face and observe it again will "be off".
Longer occupations are not for increasing precision but for increasing confidence. (Which holds up in court)
Well, as Jan Van Sickle has said a few times in the seminars I have attended [paraphrasing slightly]:
"What is the minimum number of epochs do I need to observe in RTK to get a good position?
One, if you want to be sure that no one can ever question your results..."
a USGS geodetic benchmark
I hope you mean National Geodetic Survey (NGS) bench mark, as most of the USGS marks are not of sufficient quality to be included in the NGS data base.
So the GCPs are CYA points. Efficiency wise you could do without, and usually be right.
"Usually" ain't going to cut it when performing professional services. A surgeon is not going to tell you that "I'm only going to give you one stitch after surgery, because I've noticed that the first stitch usually holds right after I sew it." ("I've noticed" is also code for "I have no hard evidence and certainly no statistics to back up my position".)
If a product or deliverable includes measurements or quantities, those numbers are useless without a confidence range. That goes for any industry or service.
I don't consider GCPs (or scan targets or deformation monitoring control, or redundant measurements) to be CYA, but rather an integral part of professional-quality work.
Efficiency means jack if you can't back up your results at some quantifiable level.
("I've noticed" is also code for "I have no hard evidence and certainly no statistics to back up my position".)
It's also an open invite to getting mauled by opposing council if you're ever dragged into a court case.
one initial observation can't be precise. it's only one observation. could be 24 satellites, multiple data points, one observation. or very well could be 10 satellites, lots of multipath, high PDOP, ETC ETC.
BUT not precise.
Yup, I mixed them up, accurate not precise.
I hope you mean National Geodetic Survey (NGS) bench mark, as most of the USGS marks are not of sufficient quality to be included in the NGS data base.
I was thinking of USCGS (Even older) but in reality the BM that crossed my mind was an NGS marker. Been a few years, it was AI9401. Usually you setup and it would hit 9.02 immediately (9.03 actual) within three minutes it usually topped off about 9.11' . Thanks for pointing that out.
2 mistakes. Teach me to post after drinking LOL
"Usually" ain't going to cut it when performing professional services. A surgeon is not going to tell you that "I'm only going to give you one stitch after surgery, because I've noticed that the first stitch usually holds right after I sew it." ("I've noticed" is also code for "I have no hard evidence and certainly no statistics to back up my position".)
If a product or deliverable includes measurements or quantities, those numbers are useless without a confidence range. That goes for any industry or service.
I don't consider GCPs (or scan targets or deformation monitoring control, or redundant measurements) to be CYA, but rather an integral part of professional-quality work.
Efficiency means jack if you can't back up your results at some quantifiable level.
Yes that was the point of my post. IT does not matter what is more accurate or more efficient. It matters what is legal. 70% confidence is not going to hold up in court.
Do you have data to back up the claim that relative accuracy and positional outliers do not improve after "a few seconds"? Because there are quite a few research papers that demonstrate the opposite.
That's like saying we shouldn't observe F1/F2 observations with a total station, because after the F1 observation if we change face and observe it again will "be off".
....
Well, as Jan Van Sickle has said a few times in the seminars I have attended [paraphrasing slightly]:
"What is the minimum number of epochs do I need to observe in RTK to get a good position?
One, if you want to be sure that no one can ever question your results..."
I like that quote 🙂
As for data to back up my claim. I am years removed from the BM I was referring to, the one we used at least once a day everytime I was in that town. I did not write it down, as I never expected to go to court over an anecdotal statement posted on an industry message board.
I was agreeing with the OP that RTK can be very accurate for his Drone work, even though they are I assume single epoch observations. However that won't fly yet. It is why we do what we do.
Someday, his research may help prove the case for pure RTK flights (along side thousands of others doing the same). GNSS has been around for a long time, and has gone from hours of a base setup on a known point and using a rover, to using a single RTK capable multiband GNSS unit needing only a few minutes of observations to get enough confidence to make it admissable evidence for court. As far as I understand it. So standards do evolve and change as technology advances.
I will try to be more accurate in my wording so there is no confusion next time 😉
Thanks BC for the video. I found it very interesting.
