Suggestions on processing settings to improve canopy reconstruction

[Malalako]

I am wondering if anyone has advice on how to improve the reconstruction of forest canopy. I know that trees are difficult to stitch together, but it’s a relatively open canopy so I’m surprised the software is struggling so much.

I flew a Mavic 3M at 70m AGL with 80% front and 80% side overlap. At this stage, I’m just processing the RGB images but will try the multispec images in case that works better. There was a bit of wind which I know doesn’t help. I’m not looking for suggestions on changing the data collection moving forward, but hoping for suggestions on how to improve the outputs with the data I already have.

Processing:
1.   Make sure image qualities are high (all over 0.8)
2.   Align cameras – accuracy setting highest and all other options default
3.   Do camera optimisation – default settings
4.   Create dense point cloud – quality setting high, all other settings default

I have played around with changing the accuracy and quality settings on the align cameras and dense point cloud stage and that hasn't really seemed to help.

Everything works ok but the canopy has been very poorly reconstructed. Attached are screenshots showing the dense point cloud produced from the Mavic 3M and also one producced over the same area using LiDAR. I understand the point cloud from photogrammetry will not be as complete as from LiDAR but was expecting something a bit better than I got.

[olihar]

Try and disable filtering and see what happens.

[Malalako]

Try and disable filtering and see what happens.

Thanks for the suggestion. I've just finished running the exame same project but disable filtering and effectively no change to the point cloud unfortunately.

[Malalako]

I've had a play around with different settings - varying the alignment quality, preselection settings, density of point cloud, and filtering (see screenshot for canopy cover % for each). Reducing the density of the point cloud had a poor effective (makes sense) and nothing else really substantially changed the results.

[Bzuco]

Guided image matching option is designed to help in areas with a lot of vegetation, so try to enable it and set Key point limit per
Mpx to ~2000-3000.

[Malalako]

Guided image matching option is designed to help in areas with a lot of vegetation, so try to enable it and set Key point limit per
Mpx to ~2000-3000.

Thanks for the suggestion. That has improved the reconstruction slightly (canopy cover [CC] up from 15.3% to 15.7% but still nowhere near a complete canopy (42.7% CC according to LiDAR flight).

The only thing that has worked is using the multispec images and changing the primary channel to NIR (results in 32.4% CC) or Red-edge (29.1% CC). Using the default (green) or red bands as the primary channel results in very poor reconstruction - 14.3% and 12.9% CC.