[GRASS-dev] How to calculate mean coordinates from big point datasets?

[Hamish]

Hamish wrote:

> see also the v.points.cog addons script:
>    http://grasswiki.osgeo.org/wiki/AddOns/GRASS_6#v.points.cog
> 
> although I haven't tried it for anything as big as lidar data.

oops, I completely forgot to mention the r.cog addon script too:

  https://trac.osgeo.org/grass/browser/grass-addons/grass6/raster/r.cog

The main goal isn't just center-of-gravity, it's also to reduce the
dataset to a single trend plane. In that way you might subtract out the
principal component of the z-surface to show the underlying signal
(or bring the primary trend into focus). This is e.g. a common task in
working with magnetic data where the Earth's field dominates, but
is relatively smooth compared to the local signal.

A good sample data for it is spearfish's elevation rasters since there
is a strong N-S trend, and for categorical features using the spearfish
fields map + choosing different combinations of categories to move the
center-point around.

I suspect that it could be quite effective with lidar if used with
r.in.{xyz,lidar}, and would be interested to see how the results compare
numerically with MMetz's new module, especially if the 10% trimmed-mean
binning method was used to solve the mean vs. median outliers problem.
(for really big datasets that method has a non-trivial RAM cost though)


r.cog is not listed on the wiki yet since the dip angle and azimuth +
r.plane trend map is still a work in progress -- help welcome, I've
got the 3D pivot point and azimuth seemingly ok, but the mean slope
has proven tricky. My last attempt at it was to split the raster into
three equal-area triangular zones (sort of looking like the South African
flag) then do the COG for each (yes, a recursive script :) to get 3 3D
points, then fit a plane formula to them. IIRC the trouble I was having
with that was if there were large areas of NULL on one side it threw
the whole thing off.


see the example in the v.points.cog help page for an idea of what its
original target application was: placing labels in just the right spot,
while r.cog was written mainly with the idea to determine/create trend-
surfaces.


regards,
Hamish