How good are classification results in practice?
    
   The following detailed commentary was posted by Chris Hermansen
   (clh@tfic.bc.ca).

Mike Joy posted a question regarding irregularities between two
classifications, one derived from manual interpretation of
large-scale aerial photography, the other from a supervised and
enhanced spectral classification of Landsat TM imagery.

I've read several of the responses, and I just thought it time
to kick in my $0.02 worth, since I am quite familiar with both
of the classifications with which Mike is working.

First, Peter Bolton rattles off his experience in tropical forests
and chastises Mike for discovering what should have been obvious.
Well, Peter, the boreal forest is a much different beast than
what you're used to in Malaysia (I can attest from firsthand
experience in both cases).  Classification from remotely sensed
data is generally quite reliable in the boreal forest, especially
given the vegetative nature of the TM-derived classification
that is Mike's second dataset.  Detecting predominantly deciduous
from predominantly coniferous stands is (spectrally speaking)
pretty straightforward.  Problems arise in mixedwood stands,
however, since the nature of the classification of proportion
is not necessarily the same and in any case any aggregative
techniques applied to the TM image prior to classification (eg
smoothing) could significantly alter the proportional balance.
Also, depending on the proportion of deciduous in a predominantly
coniferous stand, and the spatial distribution of deciduous trees
within that stand, the classifier may have difficulty detecting
the differences between mixedwood and younger pure coniferous
types.  Furthermore, deciduous stands with coniferous understory
are classified as deciduous in Mike's first dataset but may
easily be interpreted as mixedwood stands in the TM image.

Secondly, on the subject of incorporation of field data, Mike's
second dataset has some ground truthing incorporated in the
classification.

Thirdly, on the subject of large numbers of classes in some
people's TM-derived classifications, remember that in many cases
these additional classes are derived by incorporating other
datasets (field measurements, other digital map data, DEM
information, etc).  The people I've seen most test this envelope
are the folks at Pacific Meridan Resources; their TM-derived
datasets form only the first step of several.  As Vincent
Simonneaux points out, most people stop at the first step.

So, in response to Mike's original questions:

> 1) Is it reasonable to expect a TM-based classification to accurately
>    distinguish Coniferous and Deciduous forest? The area I am dealing
>    with is boreal mixedwood forest in northeren Alberta, Canada. I had
>    expected that the classification should at least be able to do this.

On the face of it, yes.  But!  You must ensure that your definition of
Coniferous and Deciduous forest is exactly the same in both cases (and
the prevailing definitions in use in Alberta don't exactly help out in
this case).

> 2) Do people out there have similar experiences, i.e. the actual
>classification
>    accuracy being very much lower than the reported results, or major
>    differences when comparing with different source of information?

Of course, this is a possibility; the most unreliable classes may
interfere in a nasty way between to datasets.  You really need to ensure
that you are sampling the same population in both cases; then you need
to examine the distribution of errors among classes in both cases.  In
your first dataset, you don't really have error estimates with which to
work.

>                                                                      I
>    understand that an air-photo-based forest inventory and a TM satellite
>image
>    are measuring different things, and that I shouldnt expect perfect
>agreement,
>    but I would have thought they could agree roughly on the overall area of
>    Coniferous or Deciduous forest. Ditto for two similar TM-based
> + classifications.

Once more, not necessarily.  See the points above on coniferous understory
in deciduous stands and the basic definitions of coniferous/deciduous
split.

There are, of course, really obvious errors that can occur, like using
pre-leaf or post-leaf images when trying to locate deciduous stands...

Sorry to go on at such length about this; I hope that my comments are of
interest to some of you.