Title: Kriging Colour

Hi,

am getting so many good replies that I thought it a good idea to compile them as I go, more to the point, before I forget.

Listed in order of receipt

Dan Cornford said:

   well with kriging of colour values then maybe it is easier simply to
remap the 0-255 to 0-1.0 and then you realise that a sigmoid function
will do exactly what you want (as we have done with classification
problems using Gaussian processes). The trick is to view the thing as a
latent variable model, with the Gaussian process being transformed by a
sigmoid function for prediction? Luckily for you the constraint is on
each component separately which makes things easier. We have recently
developed some clever sequential (online) learning methods for this sort
of problem, see:

http://www.ncrg.aston.ac.uk/cgi-bin/tr_search?logic=AND&author=csato&year=*&show_abstract=on&format=HTML

As to cokriging, well this will be of use if the colour variables are
(spatially) co-correlated! :-) One reason for preffering HSI over RBG is
that HSI tend to be less correlated, but act over different scales (good
for image compression) while RGB tend to be correlated and vary strongly
across all channels.


Vic Rudis:

Just a little, but not with kriging. We had analyzed color for vegetation analysis of digital images. What we ended up doing is subdividing Red, Green , and Blue intensity values by color into 100 'bins,' each 2.55 units wide, that is, 0.00-2.55, 2.56- 5.10,. ., 252.40-255.00. That made subsequent analysis (ANOVA, tests of means by treatment, etc.) more intuitive (0 to 100 percent of maximum intensity for each of Red, Green, and Blue)

Full report is at <http://www.srs.fs.fed.us/pubs/rpc/1999-03/rpc_99mar_31.pdf>. with image analysis methods on p. 312
--Vic Rudis <http://www2.msstate.edu/~vrudis/index.html>

Gilles Bourgault:

You can look at the paper:

"Sequential Conditional Simulation with Linear Constraints"
by Jaime Gomez-Hernandez, Geostat 2000, in Cape Town,
South Africa.

Nicholas Lewin-Koh

Isn't it more natural to think about kriging in the CIE luv or lab color
space
since they are uniform and closer to the munsel color space which is used
in the field. Interpolation in RGB space seems like a bad idea, and
could give wierd results. Also aren't there adbrupt discontinuities in
color betweensoil horizons? is color varying smoothly? 
But I think CIE LUV or LAB makes alot more sense.

Isobel Clark

Just a thought. Since your actual colour is a
combination of the three primary colours, could you
perhaps do a principal components analysis on the
three 0-255 colour levels?

This would have two effects:

(1) would give you variables which expressed the
'combined' colour

(2) would highlight the different 'populations' of
soil with markedly different colour characteristics.

We have used a similar approach to distinguish between
limestone types in quarries and for 'brightness' of
china clays.

Laura Cracker

Yes, in fact, check out this work by William Hargrove at Oak Ridge.  It
is a very effective way to visualize color data weighted by PCA results.

His paper on clustering of ecoregions:
http://www.esri.com/library/userconf/proc97/proc97/to250/pap226/p226.htm

His home page:
http://research.esd.ornl.gov/~hnw/

Laura Kracker





Benjamin Warr
Research Associate to Prof. Ayres,
Geostatistician, Environmental Scientist.

Postal Address:
Centre for the Management of Environmental Resources (CMER)
INSEAD
Boulevard de Constance,
77305 Fontainebleau Cedex,
France

Tel: 33 (0)1 60 72 4456
Fax: 33 (0)1 60 74 55 64
e-mail: [EMAIL PROTECTED]

Reply via email to