Segmentation input parameters

[Anonymous]

Hi guys,

Cool program with much needed utility! I think this will work well for a study I am conducting as part of my dissertation research, but I'm having a bit of trouble determining the optimal parameter settings for veg cloud segmentation. I tried the default settings, as well as many combinations, but my segmented tree clouds either come back as very small pieces of stem, multi-tree clusters, or a combination of both. I read through the documentation and think I have a general understanding of what the input parameters do, but when I apply the theory it doesn't always match what I am expecting. Do you have any recommendations for where to start and maybe what methodology you use to optimize these inputs to get the best initial segmentation?

Specifically, I have been working with "test data" from a single TLS scan (FARO Focus3D; ~25 million points) collected from a mostly deciduous temperate forest in Connecticut, USA. The stand is second or third growth, with maturing oak, hickory, and maple 75-125 years old. The forest is roughly stratified into understory, midstory, and overstory layers, with some branch overlap and many trees in the lower strata arcing or bending in various directions, making for a complicated digital model (see attached segmentation results).

My overall goal is to segment the vegetation cloud into individual tree clouds, then remove a subset of the tree clouds to alter the structural complexity, then cut the real forest based on the digital model. On that note, I'm also curious how you go about approaching segmentation of very large point clouds, as I will eventually need to scale up to the plot level (36 scans ~ 1.1 billion points). I assume I'll need to subset this at the very least?

Thanks!
Brandon

[Janek]

Hi Brandon,
thanks for using our software. for segmentation input parameters seems to
be the best parameters as average point distance and multiplicator depends
on data quality - shadows on branches or stem. important is also terrain
cloud, where you should remove all points that do not belong to terrain.
For very large cloud we usually use tiles made by lastools. for example we
divide 1 square kilometer inot tile with size of 25 m and do segmentation
on those tiles, with such a big data you have make a lot of manual work
when merging trees from various tiles.
As you mentioned, your forest site is temperate forest, so same type as we
have. for scanning its better to scan forest without leaves so you can
better see into the crowns.
For segmentation you can send me a piece of data and will try to find some
parameteres for you or find steps that produce good results.
Jan

po 2. 3. 2020 v 16:57 odesílatel noreply@sourceforge.net napsal:

Hi guys,

Cool program with much needed utility! I think this will work well for a
study I am conducting as part of my dissertation research, but I'm having a
bit of trouble determining the optimal parameter settings for veg cloud
segmentation. I tried the default settings, as well as many combinations,
but my segmented tree clouds either come back as very small pieces of stem,
multi-tree clusters, or a combination of both. I read through the
documentation and think I have a general understanding of what the input
parameters do, but when I apply the theory it doesn't always match what I
am expecting. Do you have any recommendations for where to start and maybe
what methodology you use to optimize these inputs to get the best initial
segmentation?

Specifically, I have been working with "test data" from a single TLS scan
(FARO Focus3D; ~25 million points) collected from a mostly deciduous
temperate forest in Connecticut, USA. The stand is second or third growth,
with maturing oak, hickory, and maple 75-125 years old. The forest is
roughly stratified into understory, midstory, and overstory layers, with
some branch overlap and many trees in the lower strata arcing or bending in
various directions, making for a complicated digital model (see attached
segmentation results).

My overall goal is to segment the vegetation cloud into individual tree
clouds, then remove a subset of the tree clouds to alter the structural
complexity, then cut the real forest based on the digital model. On that
note, I'm also curious how you go about approaching segmentation of very
large point clouds, as I will eventually need to scale up to the plot level
(36 scans ~ 1.1 billion points). I assume I'll need to subset this at the
very least?

Thanks!
Brandon

---

Segmentation input parameters
https://sourceforge.net/p/forest3d/discussion/general/thread/1ad80231df/?limit=25#a301

---

Sent from sourceforge.net because j.trochta@gmail.com is subscribed to
https://sourceforge.net/p/forest3d/discussion/general/

To unsubscribe from further messages, a project admin can change settings
at https://sourceforge.net/p/forest3d/admin/discussion/forums. Or, if
this is a mailing list, you can unsubscribe from the mailing list.

[Janek]

And one more note. From your figure it seams that you are processing only one scan from plot. It has problem with variable point density near and far from scaner. so if you have only single scan setup try to cut far away points or merge more scans together for more uniform pointcloud.
Jan

[Anonymous]

Hi Jan,

Thanks for the quick response. I think the terrain cloud I got was pretty accurate, but I see what you're saying about the single scan and leaf-on conditions being an issue. I tried running automatic segmentation on multi-scan veg clusters too, but got results similar to the single scan. Seems like most folks are using leaf-off scans for this type of thing, but our definition of structural complexity is based on explicit leaf arrangement, so I will most likely need to use leaf-on scans to meet the study objective (if feasible).

If you don't mind taking a look, I can definitely send some test data your way. I tried attaching a small point cloud to this post, but it was too big; where should I send the data, and what format would you prefer? Interested to see how well the automatic segmentation will work with appropriate input settings and leaf-on conditions.

Thanks!

[Janek]

Hi Brandon,
for data you can use my mail address j,trochta@gmail.com for data i prefer
pcd file from project folder or las file.
jan

st 4. 3. 2020 v 1:18 odesílatel noreply@sourceforge.net napsal:

Hi Jan,

Thanks for the quick response. I think the terrain cloud I got was pretty
accurate, but I see what you're saying about the single scan and leaf-on
conditions being an issue. I tried running automatic segmentation on
multi-scan veg clusters too, but got results similar to the single scan.
Seems like most folks are using leaf-off scans for this type of thing, but
our definition of structural complexity is based on explicit leaf
arrangement, so I will most likely need to use leaf-on scans to meet the
study objective (if feasible).

If you don't mind taking a look, I can definitely send some test data your
way. I tried attaching a small point cloud to this post, but it was too
big; where should I send the data, and what format would you prefer?
Interested to see how well the automatic segmentation will work with
appropriate input settings and leaf-on conditions.

Thanks!

Segmentation input parameters
https://sourceforge.net/p/forest3d/discussion/general/thread/1ad80231df/?limit=25#488d

---

Sent from sourceforge.net because j.trochta@gmail.com is subscribed to
https://sourceforge.net/p/forest3d/discussion/general/

To unsubscribe from further messages, a project admin can change settings
at https://sourceforge.net/p/forest3d/admin/discussion/forums. Or, if
this is a mailing list, you can unsubscribe from the mailing list.

[Anonymous]

Files are on their way. Hopefully they don't get routed to the junk folder!