i.cluster

[Katarina Johnsson]

In reply to:
> 
> I have been running i.group, i.cluster and i.maxlik to perform an unsupervised
> classification of avhrr images over Asia. the output from i.cluster is long and
> detailed but noone here seems to know what the numbers mean, what the units are, etc. and the documentation does not give any explanation about it either. 
> Does any one on the net knows what i.cluster produces and is there any documentation out there on how to use the result file?
> 
> *******************************************************************************
> Dominique Bachelet                *  If you plan for one year, plant rice;
> US EPA Env. Res. Lab.             *  if you plan for 10 years, plant trees;
> 200 SW 35th Street                *  if you plan for 100 years, educate the
> Corvallis OR 97333                *  people.
> Tel: 503-754-4314                 
> Fax: 503-754-4338                 *                       Chinese proverb
> Email: dominiq at rice.cor2.epa.gov  *
> *******************************************************************************
> 
Dominique,
I've been using i.cluster and there are some things I know and some things I
wonder about too. I agree, some documentation on how to interpret the results
would be quite useful. Anyhow, here's what I do know:
1) You get two output files, one is sent to you by mail and the other is
stored as a signature file. The signature file is used as input to i.maxlik
and it contains descriptions of the clusters that have been created. The format
looks like this:
#produced by i.cluster
#(here comes the description of the first cluster)
403  (A)
27.420765 49.710383 55.491803  (B)
12.893413                      (C)
6.540113 16.085991
27.368852 20.78075 74.910647

  (A) this is the number of pixels included in the cluster, it must be larger
      than the Minimum class size, that you give as an initial parameter

(B) this is the means in each band for the cluster. I'm using 3 bands in my
    classification. They are specified as a subgroup with i.group.
    I think that the order of the means in this output file from left to
    right corresponds to the subgroup as listed from top to bottom, ie the left
    most value (27.42...) corresponds to the first band in the subgroup list 
    and so on.  BUT - WARNING - I usually mix this up so you might want to
    make absolutely sure yourself. I usually check by comparing to a histogram
    or so.

(C) this is the covariance matrix (3 lines). The covariance matrix describes the
    covariance between the three bands. From the covariance matrix the eigen
    vectors can be computed and eventually you can plot the cluster in the
    3 dimensional space (I don't think you can do this with any GRASS software)
    The covariance matrix is the multi dimensional equivalent to standard
    deviation (or variance) for one image band and it describes the cluster
    statistically.
    You can normalize the 
    matrix, so that the values along the diagonal is equal to 1. The other 
    values in the matrix will then give you a direct measure on the correlation
    between any two bands in your cluster. The lower the value ( on a scale
   between 0 and 1) the lower the correlation, ie the degree to which any two 
    bands vary together. The normalized covariance matrix is also called
    a correlation matrix. I think there is a function somewhere in GRASS
    to convert from covariance to correlation matrix, but I can't remember
    where (Included in i.maxlik maybe??)

This is all the information you get for each cluster in the signature file. 
The order of the values in the covariance matrix is the same as for the means
ie the general format looks like this:

mean1    mean2    mean3
cov11
cov12    cov22
cov13    cov23    cov33

2) The file that you get with mail contains more info, basically describing
   how the clusters were generated, how they converge and so on. You also
   get a class separability matrix. I don't know how it is computed and 
   what it describes. Maybe someone else can tell? I would also like to know 
   what should be taken into account when one sets the Minimum class separation
   parameter as well as the convergence parameter?

I hope this is of some help to you! I'm not quite sure if it is possible to 
understand my explanation.

- Katarina

------------------------------------------------------------------------------
Katarina Johnsson			e-mail: kjohnsso at alder.pfc.forestry.ca
Advanced Forest Technology Lab		phone:  (604) 363-6017
Pacific Forestry Centre			fax:    (604) 363-0775
Forestry Canada
Victoria, BC