Alignment issues using PPK photos with RTK GCPs

[FredericB]

I am having some odd issues during the image alignment of a dataset consisting of 1200 photos captured with a Phantom 4 Pro V2 and KLAU PPK that I am trying to align with GCPs that were collected using a STONEX dGPS. The base station log from the GCP dGPS mission was used for the PPK processing. The mission is a double grid pattern shooting oblique imagery (75 degrees) at 75% side and 80% front overlap, since I am interested in capturing vegetation structure and house sides etc.

https://imgdrop.io/image/tHdCT

For some reason a few hundred cameras (predominantly one direction of the double grid fight path) don't align with the rest, and instead build separate planes that intersect with the real surface. In the attached screenshot it is only one extra plane but depending on how many or which GCPs I am using as control points, multiple planes can be created. I tried realigning those incorrectly aligned images but they never align with the rest of the correctly aligned images. I am getting incredibly large errors on my GCPs, even though I was hoping for <10 cm accuracy.

Any help would be appreciated!

[olihar]

Is the large error on the GCP in height? If so you have not reporjected drone images into the correct GEOid.

Can you post a screenshot showing images in 3D space before processing.

[Paulo]

Hi Frederic,

maybe show a screen capture with Workspace, aligned cameras with matching graph as well as tie points info would help understand the issue better...Is the PPK processing accurate (fixed cm level)?  If so , I would try disabling all GCPs and update transform to see if it improves...

[Paulo]

Bonjour Frederic,

how is it with your issue?

[FredericB]

Regarding geoid height, I thought I had processed everything using ellipsoid heights (and WGS84 UTM).
PPK processing in KLAU PPK resulted in high accuracy estimates (3 cm) with a few exceptions of 25 cm accuracy. Despite that I used 1 m camera accuracy settings because camera errors always end up larger.
Coordinates were uploaded to Metashape with a CSV file. Yaw, pitch and roll values are used from the original EXIF tag. I should mention though that the gimbal has a weird drift and throughout each battery during the mission, the camera yaw drifted in one direction relative to the drone yaw.

It seems like whenever I tweak any of the settings, completely different error patters appear.
Here I tried no control points (all GCPs unchecked in reference settings), unlimited tie points (which generally helps a lot with aligning this dataset). Almost all images aligned, but massive camera location errors (mean error of 96 m, range from 2 m to 400 m), and multiple meters of GCP error ( in all directions). The camera location error pattern seems like doming across the study area.

[Paulo]

Strange, Frederic.

What about the camera calibration parameters? do they look reasonable?

Maybe if you can share the alignment report, it could better understand the problem...

[FredericB]

Strange, Frederic.

What about the camera calibration parameters? do they look reasonable?

Maybe if you can share the alignment report, it could better understand the problem...

Sure here is the report. https://smallpdf.com/file#s=29c70cf0-15be-4cf9-9459-c2cdda8f9296
Adjusted camera calibration coefficients don't look too bad. k1, k2, k3, p1, p2 all <0.02, cx = 33, cy -22. This is after alignment, before camera optimization.

[FredericB]

Another thing that may effect the survey is the base station coordinates I used. I thought this workflow would be fine but now I am not sure anymore.
I am doing repeat surveys in this study area and set up the basestation over a survey marker in a concrete slap approximately 1 km away from the survey area. A while ago I collected a baselog over this marker and submitted the log to PPP and retrieved what I thought of as precise coordinates that I then used in following surveys including the one in this thread. I used these PPP coordinates as 'known base coordinates' in later dGPS surveys to collect GCP coordinates in RTK mode and also as base station coordinates for the drone images PPK processing.
But I just submitted the base log for this project to PPP and got back decimeter XY and over a meter Z errors compared to my a priori coordinates that I used for RTK and PPK.
Is this an issue? Originally I thought this would be a good procedure for comparing multiple point clouds of the same area over time, but now I am thinking maybe I should use PPP corrected coordinates for each repeat survey for PPK and to correct my RTK GCP coordinates (by adding XYZ offsets).

[Paulo]

Looking at the report, it seems there is a blunder somewhere in the input reference coordinates, either camera or GCPs. If you used same base coordinates for both GCP RTK and camera centers PPK then there should be consistency between them...

One thing I would try is gradual selection of tie points and subsequent camera optimization.

If you can share the aligned project in psz format (without images), I could have a look at it...