[GRASS-dev] grass classification routines and small / large numbers

Hi,

I have found that using the GRASS classification modules work well when the
inputs come from discreet (0-255) distributions- for example landsat
channels, etc. - however I seem to get a lot of singularity problems, or maps
with a single class when using floating point values of different magnitude.
I am imagining that this is due to scaling issues - and perhaps badly
behaving algorithms when input variables are both very small and very large.
I have found that when using clustering approaches in R, it is possible to
pre-standardize the input data, which usually results in much more
interpretable results. Are there any particular gotchas associated with the
GRASS modules which one should be wary of ?

I am mainly asking to avoid the memory limitations of R- loading 6-8 large
grids usually fills the available memory.

Cheers,

--
Dylan Beaudette
Soils and Biogeochemistry Graduate Group
University of California at Davis
530.754.7341

Dylan Beaudette wrote on 06/26/2007 07:33 PM:

Hi,

I have found that using the GRASS classification modules work well when the
inputs come from discreet (0-255) distributions- for example landsat
channels, etc. - however I seem to get a lot of singularity problems, or maps
with a single class when using floating point values of different magnitude.
  

Dylan, *,

there is i.pr available in the GRASS Addons repository from Stefano
Merler (my colleague from IRST):
https://grasssvn.itc.it/grasssvn/grassaddons/trunk/grassaddons/

"
* pr : C code for classification problems. It implements k-NN
(multiclass), classification trees (multiclass), maximum likelihood
(multiclass), Support Vector Machines (binary), bagging versions of all
the base classifiers, AdaBoost for binary trees and support vector
machines. It allows feature manipulation (normalization, principal
components,...). It also implements feature selection techniques (RFE,
E-RFE,...), statistical tests on variables, tools for resampling
(cross-validation and bootstrap) and cost-sensitive techniques for trees
and support vector machines. Feature selection techniques and
statistical tests are not distributed in the current release.
||| * i.pr : a version of pr implemented in the GIS GRASS for dealing
with images.
|"

Maybe interesting?
Markus

------------------
ITC -> dall'1 marzo 2007 Fondazione Bruno Kessler
ITC -> since 1 March 2007 Fondazione Bruno Kessler
------------------

On Thursday 28 June 2007 06:42, Markus Neteler wrote:

Dylan Beaudette wrote on 06/26/2007 07:33 PM:
> Hi,
>
> I have found that using the GRASS classification modules work well when
> the inputs come from discreet (0-255) distributions- for example landsat
> channels, etc. - however I seem to get a lot of singularity problems, or
> maps with a single class when using floating point values of different
> magnitude.

Dylan, *,

there is i.pr available in the GRASS Addons repository from Stefano
Merler (my colleague from IRST):
https://grasssvn.itc.it/grasssvn/grassaddons/trunk/grassaddons/

"
* pr : C code for classification problems. It implements k-NN
(multiclass), classification trees (multiclass), maximum likelihood
(multiclass), Support Vector Machines (binary), bagging versions of all
the base classifiers, AdaBoost for binary trees and support vector
machines. It allows feature manipulation (normalization, principal
components,...). It also implements feature selection techniques (RFE,
E-RFE,...), statistical tests on variables, tools for resampling
(cross-validation and bootstrap) and cost-sensitive techniques for trees
and support vector machines. Feature selection techniques and
statistical tests are not distributed in the current release.

||| * i.pr : a version of pr implemented in the GIS GRASS for dealing

with images.

|"

Maybe interesting?
Markus

This might be just the ticket. I was hoping to avoid a constant jump back and
forth between GRASS and R (although it may be worth it...?) -- so the i.pr
routines sound great. Any word on weather or not this module will end up in
the main distribution of GRASS ? Perhaps I can do some testing and report
back.

cheers,

dylan

--
Dylan Beaudette
Soils and Biogeochemistry Graduate Group
University of California at Davis
530.754.7341

Dylan Beaudette-2 wrote:

On Thursday 28 June 2007 06:42, Markus Neteler wrote:

Dylan Beaudette wrote on 06/26/2007 07:33 PM:
> Hi,
>
> I have found that using the GRASS classification modules work well when
> the inputs come from discreet (0-255) distributions- for example
landsat
> channels, etc. - however I seem to get a lot of singularity problems,
or
> maps with a single class when using floating point values of different
> magnitude.

Dylan, *,

there is i.pr available in the GRASS Addons repository from Stefano
Merler (my colleague from IRST):
https://grasssvn.itc.it/grasssvn/grassaddons/trunk/grassaddons/

"
* pr : C code for classification problems. It implements k-NN
(multiclass), classification trees (multiclass), maximum likelihood
(multiclass), Support Vector Machines (binary), bagging versions of all
the base classifiers, AdaBoost for binary trees and support vector
machines. It allows feature manipulation (normalization, principal
components,...). It also implements feature selection techniques (RFE,
E-RFE,...), statistical tests on variables, tools for resampling
(cross-validation and bootstrap) and cost-sensitive techniques for trees
and support vector machines. Feature selection techniques and
statistical tests are not distributed in the current release.

||| * i.pr : a version of pr implemented in the GIS GRASS for dealing

with images.

|"

Maybe interesting?
Markus

This might be just the ticket. I was hoping to avoid a constant jump back
and
forth between GRASS and R (although it may be worth it...?) -- so the i.pr
routines sound great. Any word on weather or not this module will end up
in
the main distribution of GRASS ? Perhaps I can do some testing and report
back.

Yes, would be great. After a good testing period it should go into the main
distro.

Markus
--
View this message in context: http://www.nabble.com/grass-classification-routines-and-small---large-numbers-tf3983765.html#a11356553
Sent from the Grass - Dev mailing list archive at Nabble.com.