i.pca

[Michael Shapiro]

I woke up and realized that this solution for  i.pca  won't  give
you what you need. It will only sample every 3rd pixel, that much
is true.  But the resulting output raster maps  from  i.pca  will
have the coarser resolution and this isn't good.

This memory problem will have to be fixed for 4.1.

If you have a fortran compiler, there is a workaround.   You  can
compile  m.eigensystem  (under src.contrib/CERL/misc).  This will
take an input covariance matrix and output  the  eigenvalues  and
eigen vectors. Then you can use r.mapcalc to build the components
you want. The  program  r.covar  can  be  used  to  generate  the
covariance matrix to be input to m.eigensystem. There is a manual
entry for m.eigensystem  (via  g.manual)  even  though  it  isn't
compiled.  r.covar  is  fast  and  uses little memory.  This is a
workaround (provided you have fortran) until i.pca gets fixed.

|
|One solution for i.pca is to set your resolution "larger" so that i.pca
|doesn't read each pixel, but, say, every 3rd pixel. The difference between
|the covariance matrices for all pixels and a subsample would probably
|not be significant.
|
||   "i.pca" gives me a G_malloc erros whenever I use it on an image subscene 
||   > than 900*900 pixels i.e. less than 1 Mb in GRASS compressed raster format?
||   ...I don't understand why this is so as we have a 320 H with 32Mb RAM 
||   & 64 Mb swap space (running AIX 3.2.1).