Dear Chris,
in biology CS is most of the time important (taxonomy, allometry etc.).
If that's the case in your work and you have linear measurements taken
on the same crania and views, you might be able to use those to rescale
the raw coordinates. You just need the same two landmarks between which
you measured a distance length. Those two landmarks, if not part of your
configuration, can later be discarded. You find examples in our papers
on patas monkeys and horses (those on 3D to 2D approximations): pdfs in
my webpage. In those data, I had no valid scale factor in the pics, but
had interlandmark distances (condylobasal length or the like) that I
could use to convert pixels to mm. It's a simple multiplication you can
do in xls. Once that's done, you can import the data in MorphoJ and
should have a meaningful CS.
If you used different settings for the pics (cameras, zoom etc.) there's
a possibility of a directional error. Fruciano's review on meas. error
mentions a few of examples (in those cases, it was different 3D
scanners, if I remember well).
If some of those crania are available to you, you could check by
retaking the pics with both cameras and zooms and then, maybe, applying
the meas. error ANOVA Mike and Dean published earlier this year (Evol.
Biol. this Spring). If you don't use R, a vert crude way with no test
and only preliminary and exploratory is to check how the data, including
duplicates, cluster (phenograms, ordinations): with those duplicated
data, for instance, take the means populations by camera (e.g., with 2
populations and two camera settings, you get 4 means) and check how they
cluster. Here, in Fig. 5a,
https://europeanjournaloftaxonomy.eu/index.php/ejt/article/view/2527 you
find a worrying example of bias due to small differences between two
rounds of data collection: in that case, even without duplicates, the
directional error was obvious in the ordinations. The ME ANOVA is much
better if you can do that.
As you mentioned phylogeographic differences, I guess it's mostly small
variation where the impact of a bias can be relevant.
Cheers
Andrea
On 20/10/2024 00:57, 'Chris J Conroy' via Morphmet wrote:
Thanks. I am definitely considering whether it is worth the work to rectify
the values. Most of the time I see CS it is simply a proxy for size, and I
have skull and other lengths in hand.
On Friday, October 18, 2024 at 9:48:41 PM UTC-7 Murat Maga wrote:
Before you do anything first decide whether the orders of magnitude
difference in CS is going to be relevant. Procrustes analysis will
uniformly scale the coordinates using the centroid size from each specimen,
so they all will have unit size. So, from analytical point of view there is
no harm if you are only going to use the procrustes aligned coordinates or
PC derived from them in your analysis. The moment you will run into issues
is when you include the actual centroid size as a variable in your
analysis, whether for allometric regression or for simply using it species
comparison. Imagine a case where you have two specimen of the same species,
one with centroid size in thousand (because probably the coordinates are in
raw pixel distances) versus one it is in tens (because probably same
scaling has been already applied to the image either directly or
indirectly). That obviously wouldn't make much biological sense.
If you had ruler visible in all your images, proper way to correctly scale
the data is to measure a known distance (e.g., 50mm) in ImageJ and look at
the length reported in pixels (e.g., 500 pix). Then the correct scale of
data is 10pixels/mm (or in other words 100 micron resolution), which you
can apply using the menu you have shown in your screenshot. This of course
assumes the measured distance is more or less parallel to one of the image
axes. The more off-axis it is, the more error you will make. If the field
of view and the image resolution didn't change between acquisition, then
you probably only have to do this on a few samples, and if they are all
coming out around the same value, you can use that as your the scaling
coefficient for the remaining samples without having to measure them.
However, if the field of view has changed (zoom in/out), or you changed the
resolution of the images after the acquisition you can't rely on your one
size fits all type of scaling, and have to go one by one. There are three
orders of magnitude difference in your centroid size plot, so it is hard to
tell whether the clusters around thousands range are due to different image
acquisition parameters, or natural size variation (which might be the case
if you kept all the imaging parameters the same and your specimens vary
quite a bit in size). All I can say, you definitely have minimum of two
groups for sure (thousands range, and the ones close to zero). The rest you
have to decide by trial and error.
Again, if you are not planning to use centroid size anywhere in your
analysis, then all this is moot. You can stick to procrustes coordinates
based shape analysis, and all will be good.
On Friday, October 18, 2024 at 3:40:27 PM UTC-7 Chris J Conroy wrote:
Hello morphologists,
I’ve created a landmark dataset for ~250 mammal skulls. My process
was to take images in the same view, with a ruler in the image. I opened
the files in ImageJ and plotted the points. I’ve made an input file in
MorphoJ and done the usual analyses, which largely make phylogeographic
sense. However, centroid size appears to be in semi-discreet bands of
values ranging from near 0 to near 4000. I’m sure the variation comes from
at least two sources. One is that I used two cameras, one at 72 DPI and the
other at 96 DPI. Opening the images in Mac Preview allows me to check this.
Next, in ImageJ, it looks like for some of them I set a scale based on the
image density and others I did not. The landmark XY values themselves
differ by orders of magnitude.
I took a sample of specimens from top to bottom of the centroid size
variation and fixed the scale in ImageJ and recreated the landmarks. The XY
values are now in the same range.
Before I sit down and do this for all 250 images, I wanted to see if this
is going to be a waste of time, or will images from different digital
cameras going to be fundamentally different in centroid size and not worth
the work. I suppose I could do the work and then code all the images by
camera and see if there is an obvious effect. Surely I’m not the first
person to go through this.
If it comes to it, I can re-take the lower res images on the higher
resolution camera.
Lastly, on the higher resolution camera, some images are slightly zoomed
to fill the image with the object. Will this also mess up centroid size?
Thanks,
Chris
--
Andrea Cardini
E-mail address: [email protected], [email protected]
WEBPAGE: https://sites.google.com/view/alcardini/
--
You received this message because you are subscribed to the Google Groups
"Morphmet" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
To view this discussion on the web visit
https://groups.google.com/d/msgid/morphmet2/d7c10d4a-86d1-4d53-a024-447c8c6d318e%40gmail.com.
