I am returning to the ground classification issue once again (recently I posted similar issue on subterrain noise). I went through all of the posted and recommended tutorials on classification of noisy mostly photogrammetry data but none gave satisfactory result for me (maybe I do not understand the system of percentile of elevation enough, which may be a cucial part of the whole procedure) .
In most parts of the landscape I am ok with my workflow but there are significant parts where I fail.
My workflow is as follows:
2. Lasnoising /xy step 0.8, z step 0.2, isolated points 8/
3. Lasground /step 0.8, spike up 0.35, spike down 0.02, offset 0.02, bulge 0.02, hyperfine/
4. Statistical outlier removal /SOR/ (performed in CloudCompare) on ground points
Agressive lasnoise parametres are able to catch majority of undreground noise and eat some parts of the vegetation as well which I dont mind as I am interested purely on ground. However the are still underground noise points left.
Lasground with the given parameters works fine on approx 70 perc, of the landscape, but in some parts it misclassifies remained underground noise as ground and real ground above it as unclassed.
In most cases I was able to fix the problem with the final step - SOR performed on "grounds only" pointcloud, which helped me to ged rid off most of remained underground noise and remains of vegetation that lasground was not able to fix.
But currently I struggle with a lot of misclassified real ground which is omitted (misclassified as class 1) and only parts of the subterrain noise undrneath is left (misclassified as ground). In this case final SOR only produces holes as it gets off underground points (with lower density than the average) but exact ground points above are missing.
I would appreciate any comment
PS: explanation to pics: red -noise point
brown - ground
PS2: As far as I know the data was collected with a RIEGL "crossfire" LMS Q1560.