[Solved] Trouble with GCP marker error UAV survey

[obw]

Hi there.

I've been having some trouble with my latest UAV survey work where I've been getting high marker errors on most of my GCP's. I've been searching through the community and found a lot of useful tricks and have optimized my workflow, but unfortunately I got the same results with high errors. The latest attempt was With SAV's optimized workflow (http://www.agisoft.com/forum/index.php?topic=7851.msg37494#msg37494) for UAV survey, which is pretty similar to my own workflow that I've created from your tutorials and community reading.

Aircraft in use:
DJI Inspire 2 X5S 15mm lens. 954 images.

GCS in use:
Leica GS08 Antenna with CS10 hanheld unit, survey grade accuracy. 23 GCP placed.

I've been doing some other projects with this setup without any problems.

It seems like the biggest problem is that the altitude from the DJI photos is not correct. We flown aprox 80m, but the image description says negative 2-4m. But the photos still allign fine so I guess this sould not affect the accuracy as long as we're using GCP? But still, when I add the GCP markers, the errors I get is aprox the same as the diffrence in height (70-75m in error). When I linked the markers to the photos and "updated" the error gets down to a couple of meters, when I hit "optimize" (with standard parameters) the error gets lowered to more normal values, between 3-25cm, but unfortunately most of them above 5cm which is too high. The photos are pretty good, no problems with the allignment and I've removed the bad quality photos.

I'm attaching some screenshots and the report generated.

Thanks for all your help, please let me know if you require more information.

[obw]

Report Attached, had to compress it to upload it

[JMR]

Try fit rolling shutter, the X5S camera is BAD for photogrammetry. If you want to get much better results, buy an X4s or at least fly sloooow or even better if you stop at every photo shot.
Best regards
Please tell us if results get better.

Regards,
GEOBIT

[cadm8]

Also some of your erroneous GCPs are on the edge of the project, so maybe too few projections for them? Try re-arranging control points and check points. Still though, is there a possibility to have errors in your coords from the GPS? Did you survey by RTK, maybe you have one or two rogue observations? Finally why not using 40k and 10k for key point and tie point limit?

Personally I'm suspecting that you have a couple or more wrong GCP coords, try selecting a minimum number of GCPs and then adding slowly the rest until you spot the problematic ones.

[obw]

Try fit rolling shutter, the X5S camera is BAD for photogrammetry. If you want to get much better results, buy an X4s or at least fly sloooow or even better if you stop at every photo shot.
Best regards
Please tell us if results get better.

Regards,
GEOBIT

Interesting, can you elaborate why this is the case with the X5S? We recently upgraded from an Inspire 1 with X5R and are satisfied with the raised image quality this gets us. Are the X4S better than both of theese in that matter? When that is said, we never had accuracy issues with the X5R...

Also some of your erroneous GCPs are on the edge of the project, so maybe too few projections for them? Try re-arranging control points and check points. Still though, is there a possibility to have errors in your coords from the GPS? Did you survey by RTK, maybe you have one or two rogue observations? Finally why not using 40k and 10k for key point and tie point limit?

Personally I'm suspecting that you have a couple or more wrong GCP coords, try selecting a minimum number of GCPs and then adding slowly the rest until you spot the problematic ones.

I've checked all my GCP on the CS10 unit, and they all had at least 5 recorded RTK positions and all with maximum error of 0.0025, while most of them was below this.

I'll try to remove some of them, but do I have to reload the project before every try, or can i just uncheck all - and check one by one and optimize between every one?

I first alligned the photos at High with normal point limits, but raised it up to highest and raised the tie point limits to try to improve accuracy, but the results were the same.

[obw]

Try fit rolling shutter, the X5S camera is BAD for photogrammetry. If you want to get much better results, buy an X4s or at least fly sloooow or even better if you stop at every photo shot.
Best regards
Please tell us if results get better.

Regards,
GEOBIT

This was the answer! I did not notice that the rolling shutter option had moved to the camera optimization tab in version 1.4.0, so it was unchecked. When enabled, my average error dropped down to an average of 0.037 which is within limits. Thank you!

[JMR]

Interesting, can you elaborate why this is the case with the X5S? We recently upgraded from an Inspire 1 with X5R and are satisfied with the raised image quality this gets us. Are the X4S better than both of theese in that matter? When that is said, we never had accuracy issues with the X5R...

You are welcome!. Rolling shutter is killing the quality of many surveys, and even worse is the fact that too many people close eyes and swallow while they deliver bull shit. The otherways lovely X5 is an electronic shutter camera that has a terribly slow sensor readout. That, in fact, means that your photos are not taken from one point, but from a sucession of as many points as rows are in your 20mpix photos one after one. The faster you fly, the longer will be the "trail of points of view". So, if PS calculates camera coordinates as one single point... what happens? Far from having an easy answer under the classic frame photogrammetry maths. The best it can do is to counteract the deformations on the image that rolling produces and suppose that the line of points of view of an image, can be simplifyed as a theorical single point... which is not strictly real.
rolling shutter compensation is a patch or a "trick" that improves quality reasonably up to a certain level. But dont ever expect the same results you could pre-estimate if you had a global shutter camera and kept all the other factors the same (factors like focal lenght, sensor size, ground structure, flight altitude, stereo geometry, etc)
That's why you can expect better accuracy from the often underestimated in comparison X4s, which uses a true shutter even if is not a perfect global shutter.

[cadm8]

I've checked all my GCP on the CS10 unit, and they all had at least 5 recorded RTK positions and all with maximum error of 0.0025, while most of them was below this.

Surely your unit will have them logged as accurate, although this accuracy level is unrealistic and... it doesn't exist. Uncheck them all but a few with lower readings and start adding one by one. You may even have more than one outlier in your project. Also, I believe that the rolling shutter algorithm is producing over optimistic estimations in accuracy, despite the fact that is actually working correctly. We have had older projects that have been flown with a phantom 3 pro (rolling shutter) and older versions of photoscan (no rolling shutter compensation) and cross checks with total station and GNSS data have shown fabulous results.

[SAV]

Hi obw,

The rolling shutter issue has previously been discussed in detail. There might be some useful info for you in this thread: http://www.agisoft.com/forum/index.php?topic=7653.0

As already mentioned by JMR, the X5S is not the best choice when it comes to photogrammetry. The X4S is better due to the mechanical (leaf) shutter. Note that the X4S camera is basically the same camera as used on the DJI Phantom 4 Pro and Phantom 4 Advanced, hence there is not really an advantage of buying and Inspire 2 with X4S over a Phantom 4 Pro (unless you also use it for cinematography, in which case the Inspire 2 has some advantages). Additionally, flight time is better on the Phantom 4 Pro compared.

Altitude information in EXIF from DJI imagery is problematic/incorrect, hence you should uncheck all images in the reference pane and only use your highly accurate ground control points. Alternatively, you could adjust the camera accuracy (default is 10m) to a much higher value to 'deal' with the large error.

Hope that works for you.

Regards,
SAV