Consistent Inconsistencies in DSM/DTM from UAV NADIR missions

[rtfraser86]

Hello All

I am consistently (across multiple projects for months) getting this effect occurring in the DSM/DTM generated by Agisoft when using drone photogrammetry with principally NADIR photographs (See attached photograph).

In this instance, a relatively flat area has been mis-aligned and has an area that is approximately 1m higher than the rest.
Taking oblique photographs while doing the UAV survey does not fix the issue - in-fact, it makes this more common-place.

I was wondering if there was a simple, easy, fix for this.  I have spent dozens or hundreds of hours adding control points and then re-optimising the alignment, dense-cloud computation and DEM generation, as this effect is only evident once the DEM has been created, and is missed in the dense point cloud.


Settings commonly used are:
- Alignment - High Accuracy, Source reference (PPK UAV), 40,000 key point limit, 4,000 tie point limit, fixed focal length of camera
- Several times I optimise the alignment, adding GCPs and/or control points as necessary.  I also use "Gradual Selection" methods to remove erroneous points.
- Dense Cloud - High Aggressive Filtering
- DSM generation from Dense Cloud

- UAV - Mavic 2 Pro with third party PPK attachment (yes, I am aware of rolling shutter issues)
- Flight Planning - DroneDeploy iOS app

Issue occurs even when I use 100,000 key point limit and 10,000 tie point limit for photos.

[rtfraser86]

Some more examples from a totally unrelated project

[Alexey Pasumansky]

Hello rtfraser86,

Would it be possible to share any original image set that you have problems with and the project file in PSZ format with the alignment results? If you can share the problematic dataset, please send the download link to support@agisoft.com, so that we could study the issue on our side.

[mhnewton]

Hi,

Has there been a solution found to this issue? I'm experiencing something similar and my temporary workaround was to remove low confidence points (0-10). This seems to work but the resultant DEM was still noticed to be off of the actual GCP elevations, which may be part of a larger, unrelated model accuracy issue. I would have assumed that if I use a GCP as a control point that the point cloud would force itself through the GCP?
Model accuracy is very important for the project I am working on as I need to compare 2 models of different time periods to measure erosion and deposition over time.

any help is appreciated!

[Alexey Pasumansky]

Hello mhnewton,

Can you share the sample dataset, so that we could check it on our side?

Usually such artifacts, as shown on the screenshots above, indicate that there are alignment problems and subsets of overlapping images for some reason do not have almost any common tie points.