based on the wiki for Principal Components Analysis [0], numerous
discussions in the mailing lists [1,2,3,4], particularly a comment by
Edzer Pebesma [5], and personal demand, I have fixed a few issues in
i.pca in trunk r49090.
- the faulty or missing centering of the input bands described in [0]
for should be fixed
- i.pca has a new flag -n to normalize input bands with (x - mean) / stddev
- values of the output maps are now calculated depending on the input
band transformation (centering or normalization). Is this OK?
- Eigen values, (vectors), and [percent importance] are now written to
stdout instead of stderr
The results of i.pca for the examples using SPOT imagery in the wiki
[0] are now identical to R's princomp() results. If the new -n flag is
used, the results of i.pca are identical to princomp(center = TRUE,
scale = TRUE).
Tested also with 9 input maps in a region with 400 million cells.
Sounds like a very sensible couple of fixes to me.
Could these be backported to 6.4.x?
Cheers,
Ben
On 11/04/2011 11:53 AM, Markus Metz wrote:
Hi all,
based on the wiki for Principal Components Analysis [0], numerous
discussions in the mailing lists [1,2,3,4], particularly a comment by
Edzer Pebesma [5], and personal demand, I have fixed a few issues in
i.pca in trunk r49090.
- the faulty or missing centering of the input bands described in [0]
for should be fixed
- i.pca has a new flag -n to normalize input bands with (x - mean) / stddev
- values of the output maps are now calculated depending on the input
band transformation (centering or normalization). Is this OK?
- Eigen values, (vectors), and [percent importance] are now written to
stdout instead of stderr
The results of i.pca for the examples using SPOT imagery in the wiki
[0] are now identical to R's princomp() results. If the new -n flag is
used, the results of i.pca are identical to princomp(center = TRUE,
scale = TRUE).
Tested also with 9 input maps in a region with 400 million cells.