rayextract with time series data

[ZaneTCooper]

Hello,

I'm working with time series point cloud data right now and searching for solutions to minimize the work of segmenting the same tree multiple times. The automated segmentation from rayextract can vary enough over multiple years that it still requires manual correction to ensure we are capturing change from actual structural changes instead of changes in the automated segmentation.

Currently, we use a very coarse script that searches for all nearby points in a similar point cloud to try to minimize the amount of manual correction, and I think you could implement something similar with raysplit. I'm thinking of a different solution that would use the tree locations from a single time stamp to 'assist' with segmentation in other time stamps. One avenue for this may be using the terrain mesh generated from an individual tree as the input for the plot level rayextract trees, hopefully leading to the segmentation of mostly the same tree. I think this is far from the intended use, so my hopes aren't very high.

Do you have a better idea for how I could put something like this together? Most of my ideas stem from the same concept of using the 'seed points' of a tree from one time stamp as input data when processing other time stamps, but I don't know of any built in option for this.

Thanks,
Zane

[tim-devereux]

Hey Zane,

We are doing a similar thing at UQ. I've written this code for masking between time series. I think this is similar to your existing method though, and does require manual correction if there is significant change.

https://github.com/Leaf2Landscape/point-masking

I agree a more sophisticated region growing method would be useful, and is something we would be keen on too.

[thomas-lowe]

Another option may be to use the feature/alpha_weighting branch, it weights path connections based on the point alphas. You could then find the transition surfaces between the previous segmentation colours and softly downweight the corresponding points in the new scan to encourage tree segmentations to be similar.

Or upweight new points that are inside the cylindrical representation of the old extracted trees.

[ZaneTCooper]

Thanks both for the replies!
@tim-devereux I took a quick look but its hard to tell if this would be more accurate than our current method. Even if its faster, the bottleneck isn't with processing speed but with accuracy between time stamps. I'll keep it in mind if we have anyone else doing a lot of masking between time stamps soon.

@thomas-lowe Thanks for pointing out the branch, we hadn't noticed that before. Wouter and I were looking through it, but it was still hard to tell how exactly the model outputs will change relative to the alpha values without some experimentation. I'm hoping that I can carve out a day or two to mess around with it soon while continuing processing. Maybe a paired method that limits where the terrain meshes are and uses point alphas could really funnel it into segmenting the same trees. If I get to trying it out I'll let you know.

Just out of curiosity, as the alpha_weighting is an optional parameter in rayextract, why did this end up being a separate branch? Do you think it would get merged into main at any point?

Thanks,
Zane