Trees!

[holup]

Hi!

I'm testing Photoscan for mapping purposes and the goal for now is to get an accurate DEM, in order to extract important features (road edges, buildings, etc) and also obtain ground contours.
My test area has plenty of trees and I wanted to know if anyone has experience on removing this "noise" from models / point cloud, as I require only the surface for the contours.
I wanted do know if anyone has experience with this (apart from masking every photo) Any pointers?

Thanks in advance.

[George]

Hi Holup,
I beleive that this is practically hugely impossible task. For this purposes National Land Survey service of different countries use LIDAR in order to get ground level precisely.
However, it may vary from project to project in your case. For dense forests it should not work definitely. Though you must have collected GCPs at the ground level at as many places of your site as possible then deleting the rest of the point cloud.
Photoscan is a pefect tool especially for DEM/DSM purposes at which LIDAR is either hugely expensive or little bit coarse as practice shows.
Or ... am I wrong?

[George]

just check this ...

lasheight and lasground at http://www.cs.unc.edu/~isenburg/lastools/
can do a job of extracting ground level, however, based on LIDAR scanning data ...
Though the question is how one can do it based on the aerial photography point cloud which is not a LIDAR?

Finally, read this answer from LASTOOLs developer
https://groups.google.com/forum/?fromgroups#!searchin/lastools/photography/lastools/9xsy9gSAw10/wPvfH7YR2N8J

"... lasground was not
designed to work with photogrammetric data but seems to work fairly
well in practice. but if you have no "ground returns" in a dense
forest i don;t see how you can remove the vegetation ..."

If you do have any hint of a ground at your imagery ... go for it and, please share your results ... would be interesting to learn it ...

[holup]

George, thanks for the feedback. I will look what LAStools can do, it seems interesting.
Another possible approach I've tested briefly is to select data using the RGB value, it's not perfect and a lot of "segmentation" of the data would have to be done, but it can work (sort of)

[Herman]

Hi

I have experimented with the LASTools and the results looks very promising.  Trees, and buildings where very successfully removed. It is semi desert terrain and a mining operation , but still very
impressed with the ease and the speed.
I only struggle with a huge shift in the LAS data, this seems to originate in PS , not in LASTools.
Would appreciate any possible explanation for this

Regards
Herman

[George]

Hi Herman,
the question may come for densely forested areas where one can hardly capture the ground level from above (with a photo camera).
However, I agree that if your aerial imagery contains sufficient amounts of ground spots then, probably, LASTools can help
George

[Santiago]

Hi

I have experimented with the LASTools and the results looks very promising.  Trees, and buildings where very successfully removed. It is semi desert terrain and a mining operation , but still very
impressed with the ease and the speed.
I only struggle with a huge shift in the LAS data, this seems to originate in PS , not in LASTools.
Would appreciate any possible explanation for this

Regards
Herman

Hi Herman,

Could you please explain how did you manage to do it. I tried it by exporting the points in a .las format from Agisoft, afterwards I used ArcGIS using the lastoolbox that came with the las tools file. I used lasground to create another .las file that had vegetation correction but when the process was over it didn't create anything. Do you know what I'm doing wrong? Did you use ArcGIS to do the processing? Do you have the free version or the licensed one?

This is a very interesting tool, I would be using it very often and off course would be willing to pay for it, but not before I know it does work.

Best Regards and thanks for any feedback you can give me,

Santiago

[Herman]

Hello Santiago

I used the las data exported from Agisoft and used the "lasground" function. Data was then used in GlobalMapper. Ver14 of GM has the possibility to only select the "ground" points (2) classified as such by the lasground function. The "shift" I have experienced before  was not caused by Agisoft (appologies), but by GM. It however seems that ver14 has solved this - just don't use the "South Orientated" option in TM projection. I have now done some projects in more densely forested area's now and here lasground was not that successful. Terrain was mostly covered in 1m high grass and otherwise with very dense trees 2-3m high. No clean area clear of vegetation to start off with. I am sure with a better understanding of the lastools I would be able to do better.   In the end a day of visual editing (300ha) with GM was sufficient to remove the "tree points" and leave sufficient ground points . This was verified with some spot shots taken and tested by comparing with the interpolated ground surface. I barely know the lastools and hope that with more understanding to be more efficient. As mentioned by George Lidar is the answer , also with the more "intelligent" classified data (first, last, soft , hard return data) of Lidar lastools would work much better.
As long as you can see the ground and have the ability to mass edit your point cloud with the image as a backdrop you can still achieve results.
Regards
Herman

[Santiago]

Hi Herman,

Thanks for your reply. I know what I was doing wrong now, I was exporting the .las file in WGS84 and it had to be exported in a projected system (meters not degrees). Still after processing the data I wasn't getting any good results so I tried talking with Martin Isenburg, the creator of lastools. He asked me to send him the file I was creating and afterwards he answered this on a thread. If you guys are interested you can look at it here: https://groups.google.com/forum/?fromgroups=#!topic/lastools/oK81EEr7W3Y

Also it seems global mapper can handle some of that data much better than ArcGIS, I'll give it a shot.

Regards,

Santiago

[jondandois]

There are other terrain filteirng algorithms / packages as well, for example ALDPAT (http://lidar.ihrc.fiu.edu/lidartool.html) and MCC-LIDAR (http://sourceforge.net/projects/mcclidar/)

We have done a lot of terrain filtering for getting ground under trees to estimate tree height and I agree it is definitely tricky.  One thing to note, LASTOOLS is not free and I'm pretty sure that some of the tools add jitter noise to your data.

From your original picture, it looks like you have plantation plantings which means you should have decent view of the ground between the rows.

One of the biggest problems we have is the blunder noise points in the raw computer vision clouds.  These are ones that are 10's to 100's + meters from any real value and tend to really screw up the DTM filters.  Many people will just manually remove and such bad points, we try statistical filtering to avoid manual work.  One trick that has worked well for us is to calculate something like the median height value of points within a uniform grid (like 1m x 1m ) then give that significantly smaller dataset to the terrain filtering algorithm.  This seems to work pretty well.

[RalfH]

Yes, gridding the data and applying a median filter works well in my experience.