While I am by no means the voice of authority on these matters, I do have a
certain amount of experience, having grappled with the same considerations
in our study of the Cape Sable Seaside Sparrow (see references below).
First of all, let me congratulate you on having got hold of Ikonos
data. Like Phil Nott, I would like to know how you did that without
leaving behind an arm and/or a leg.
Secondly, the matter of spatial autocorrelation is simply unavoidable in
almost any ecological research. Good habitat is almost always spatially
autocorrelated; that is, if an ecological event takes place in a location
with certain characteristics, there is usually a high probability nearby
locations possess the same or highly similar characteristics. In our case,
we usually found sparrows in pixels surrounded by pixels with the same
spectral, vegetative and hydrological properties. In the case of
validating a signature, I don't think it is a problem as the choice of a
signature is mainly a spectral problem, not a spatial problem.
I think the way you collected your training samples will stand up to
scrutiny. I do think you need to be careful in properly locating your
sample points, both in space and on the imagery. Using such high
resolution imagery means you need location data accurate to within about 10
cm. This can only be accomplished using differential corrections to your
GPS readings. Similarly, I would not trust the rectification of the
imagery as it comes to you. I suggest you do a very careful
re-rectification using carefully GPSed locations of your ground control
points. Again, you need to get your rms error down to centimeters or
below. Otherwise, you will not be sure that a pixel on the image
corresponds to the tree located on the ground.
I infer from your message that the location data comes from USFS data. In
other words, somebody else did it. I would give their methodology a very
careful look and then resurvey a representative sample in order to satisfy
myself that the data are accurate to the precision you need. I find that
ecologists seldom understand the limitations of GPS data. They tend to
believe the claims that are written on the box. Careful calibrations are
absolutely necessary.
I think there is no question that partitioning your data into a training
set and an evaluation set is legitimate. In fact, you are fortunate in
having so much data to work with. It opens up a great many possibilities
for additional analysis. You can partition the data randomly a number of
ways, use one partition as the training sample and see how well the
classifier does on the other partition. I suspect you will find that there
will not be much difference from one replication to the next. If there is,
you will know that there is something weird going on and you can do a
deeper investigation.
I further infer that you are doing an exhaustive species map. If not, you
need to be careful in using the maximum likelihood decision rule as it
tends to classify every pixel as one thing or another. Also, as a matter
of personal interest, I would like to know what you used for your
priors. A reply off list would be appreciated.
Cheers,
Bob
References:
2003. Jenkins, C.L., R.D. Powell and S.L. Pimm. Demonstrating the
destruction of the habitat of the Cape Sable Seaside Sparrow. Animal
Conservation, 6:29-38
2003. Jenkins, C.L., R.D. Powell and S.L. Pimm. Why sparrow distributions
do not match model predictions. Animal Conservation, 6:39-46
Robert D. Powell
Congress Farm Research Institute
Wilmington, Ohio, USA
bob_powell@erinet.com
Ludere cum sacris
This archive was generated by hypermail 2b29 : Thu May 06 2004 - 07:05:01 PDT