We've done an unsupervised classification
and Ordination analysis of a Multispectral
image over a sub-humid Mediterranean forest.
First, we have used a segmentation technique
to tile the image into patches, and have
extracted the statistics of each patch
into a Multivariate table. From then on,
our individuals are the patches (not the pixels)
and our descriptors are the means in each
spectral band.
Second, we have submitted a random
sample (3000 patches) of the table to a
model-based hierarchical clustering and
defined a threshold to cut the tree.
Third, we have used the centroids to define
a canonical analysis, have
classified all the table in this canonical space,
have created a classified image using the
class for each patch
and have used the classified image and the plot on the first
2 can axes to "understand" the vegetation pattern
in the terrain.
Fourth, we have identified clear
vegetation units in the terrain and have projected them
on the can plot.
With this information, we have defined new centroids for
clear vegetation types (so now we have a supervised
approach), and have run a second canonical analysis.
Note that in our approach we want to be conservative
in the number of classes: we want to define clear
classes and we assume that a part of the image will be
somehow intermediate between 2 or more classes.
The second can. transformation is not too different from the initial
one, but now we have an important part of the image
with low likelihood values to any centroid. These patches
are spatially contiguous, and consolidate into few large
units. This result indicates
that we should define more centroids: in addition to
the "pure" and "clear" forest types that we ahve been able
to recognise, there is probably at least one more "mixture" class.
(This makes sense considering the history of past disturbances
and management in the region).
Well, the question is that we want now to add an unsupervised
centroid (or more than one, eventually) to the list of
supervised centroids that we have. We could just mask the part
of the image that is well classified (in terms of a high
max. likelihood) and just classify the rest, but then we could have
some new centroids that would be too close to the ones that we
already have.
We would like an algorithm in which we could input the list of centroids
that we have, and then have the algorithm defining newer ones for
those individuals that are "too far" from any in the input
list.
Any idea of such an algorithm?
Thanks
Agus
Dr. Agustin Lobo
Instituto de Ciencias de la Tierra (CSIC)
Lluis Sole Sabaris s/n
08028 Barcelona SPAIN
tel 34 93409 5410
fax 34 93411 0012
alobo@ija.csic.es
http://pangea.ija.csic.es/alobo
This archive was generated by hypermail 2b30 : Sat Feb 10 2001 - 10:36:01 MET