application if data is from different years

[wiesehahn]

Hey JR,
I just tested lidRalignment and I am amazed it seems to work quite well for the plots I tested.
I have a bunch of MLS pointclouds I want to align to ALS data, some of them have no georeference at all, and it still seems to work for the few I tested so far (although I havent checked in detail).
I am wondering whether I should also apply the alignment on the plots which have already some georeference. The problem is that ALS data and MLS data are from different points in time and individual trees might not stand anymore in MLS data. Do you think your processing can handle cases where a small amount of trees (and thus the CHM) differs between ALS and MLS data (or could the existing georeference get much worse)?
If I understand it correctly you use "only the 50% of point pairs with the smallest distances" to compute RMSE, so a minor amount of trees might not be a problem?

ps: you might want to update this to use PTD for ground classification instead of CSF

[Jean-Romain]

I have already aligned scenes where a significant portion of the two scenes differed (total cut vs trees). Of course, there are limits to what can be handled. You must have matching features. If only one or two trees are missing, this is usually not a major issue. It largely depends on the type of scene.

In coniferous forests, the canopy structure often provides very strong shape matching, so a few missing trees generally do not affect the alignment significantly. In more homogeneous deciduous forests, especially when there is little variation in CHM height, the alignment can be more challenging.

Overall, the pipeline was designed with robustness as the primary objective rather than maximum accuracy. This means that when two scenes are already very similar (properly aligned north orientation, same point of view, and high-quality data) an ICP workflow in CloudCompare may achieve better accuracy.

However, when the scenes differ substantially (for example with large angular errors (such as north/south inversion), different acquisition viewpoints (ALS vs TLS), or different data qualities (e.g. a poor one scan TLS versus a high-quality MLS scan), lidRalignment provides a highly robust solution.

[Jean-Romain]

ps: you might want to update this to use PTD for ground classification instead of CSF

Yes I must do it

[wiesehahn]

this sounds good, I will give it a with larger radius over the weekend.

[wiesehahn]

a follow-up question:
i would like to also have ground classified in the referenced MLS data, this is always done as part of the processing (to have it comparable) but not stored in the final referenced data. I guess the reasons are that this process focuses on alignment only and also because we classify ground for the circular subsample only anyway. Now I could transform the data with transform_las and do a ground classification seperately but this would read/write two times. Can I also use the matrix in lasR to do something like this and read/write only once?

library(lidRalignment)
alignment <- AlignmentScene$new(als_file, mls_file)
alignment$align()
M <- alignment$get_registration_matrix()
crs_als <- sf::st_crs(readLASheader(als_file))

crs = set_crs(crs_als)
mat = load_matrix(M)
trans = transform_with(mat)
ground = classify_with_ptd()
write = write_las()
pipeline = crs + mat + trans + ground + write

[Jean-Romain]

The lasR pipeline used internally to reproject the full points cloud is

    mat = lasR::load_matrix(M)
    reg = lasR::transform_with(mat)
    crs = lasR::set_crs(crs$wkt)
    io  = lasR::write_las(ofile)
    pipeline = mat + reg + crs + io
    ans = lasR::exec(pipeline, on = las, progress = TRUE)

So you are correct with your pipeline. You can insert a ptd. Be careful, the original pipeline is streamable meaning it can re project all your points even if they do not fit in memory. If you add a ptd stage it becomes non streamable. That may or may not matter for you.

Other option is to classify on your side your point cloud before the alignment. In this case lidRalignment will use pre-classified ground points.

[Jean-Romain]

Not related to your question but are you working with TLS/MLS? In this case I may have something interesting for you. Send me an private email.

[wiesehahn]

Other option is to classify on your side your point cloud before the alignment. In this case lidRalignment will use pre-classified ground points.

In Description of the internal pipeline it states:

For this reason, we do not rely on any pre-computed ground classification that may be present in the input files.

So I thought this would not work.
Anyway, non-streamable is fine for me I guess.

[wiesehahn]

Not related to your question but are you working with TLS/MLS? In this case I may have something interesting for you. Send me an private email.

Always interested in things you work on, but TLS/MLS will just be a minor side project for now. I like large scale applicability of ALS way more :)

[Jean-Romain]

In Description of the internal pipeline it states:

For this reason, we do not rely on any pre-computed ground classification that may be present in the input files.

So I thought this would not work.

Correct! The doc must be updated. I changed my mind when we came with datasets with very sparse ground points that are impossible to classify when you only have only 400 m². But the original classification performed on 1 km² was good. So now we are relying on the original classification (if available) for two reasons:

• We can assume it is usually good especially ALS
• The pipeline relies less on the ground than before when I wrote the doc.

The doc must be updated.