[GRASS-user] strange behavior of i.cluster with sample=1,1

[Sylvain Maillard]

dear grass users,

i'm working on landsat imagery and i need to do some automatic
classification so i'm doing several tests with i.cluster (first, after that
i will also try i.cass or i.gensigset). As I wanted to see the impact of the
"sample" parameter, i launch a script with several variations and i noticed
this result: when the sample parameter is becoming smaller, i.cluster tend
to converge to only 1 class ...
I was thinking that with more samples the result will be more accurate but
it seems to be the oposit ... with the default value for sample, the result
is always correct ...

Is there something wrong with the fact to use sample=1,1 (or sample=10,10)
instead of the default value sample=56,71 ?
for the moment i'm working with landsat 5 or 7 imagery, so in the case of
land use with a very reduce location (ex: "small" rivers), having samples
only each 56x15m=840m can occult some land use category ...

I have the same error on  gras 6.3 or 6.4, grass did'nt notice any error,
and I have enought of ram (only 65% used) ... do you have an idea of what is
going wrong and what can i try to solve this problem?


I join the results of i.cluster with 20 classes and sample=1,1 parameter
set, the others have default values => it goes wrong imediately ... and with
default sample (56,71), everything works ...

Paramètres du cluster
  Nombre de classes initiales:     20
Taille minimum de classe :        17
Séparation minimum entre classes:     0.000000
Pourcentage de convergence:           98.000000
  Nombre maximum d'itérations: 30

Intervalle d'échantillonnage en ligne:         1
Intervalle d'échantillonnage en colonne:         1

Taille de l'échantillon : 40439680 points

[ ...]

Distribution des classes
    1826562      24892     450783     312030     852164
    1646111    2323198    2780468    3019891    3214292
    3472239    3559417    3246739    2739526    2249348
    2013603    1977504    1754731    1225898    1750284

######## itération 1 ###########
2 classes, 0.01% points stables
Distribution des classes
   38277534          0          0          0          0
          0          0          0          0          0
          0    2162146          0          0          0
          0          0          0          0          0

######## itération 2 ###########
1 classes, 5.35% points stables
Distribution des classes
          0          0          0          0          0
          0          0          0          0          0
          0   40439680          0          0          0
          0          0          0          0          0

######## itération 3 ###########
1 classes, 100.00% points stables
Distribution des classes
          0          0          0          0          0
          0          0          0          0          0
          0   40439680          0          0          0
          0          0          0          0          0



-- 
Sylvain Maillard
Doctorant en Sciences de l'Environnement

Laboratoire Chimie Provence - UMR 6264 / Université de Provence

la Tour du Valat - Centre de recherche pour la conservation des zones
humides méditerranéennes
Le Sambuc
13200 Arles
France
mail: maillard at tourduvalat.org




tél:04.90.97.29.79
fax:04.90.97.20.19
www.tourduvalat.org
-------------- section suivante --------------
Une pi?ce jointe HTML a ?t? enlev?e...
URL: http://lists.osgeo.org/pipermail/grass-user/attachments/20090924/5c5b2c0f/attachment.html