— A ANI PR
um RENI tt
nc nl RENE i
- 1492 -
IV Supervised classificat ion
Using the described method we have overlaid 5 Landsat subframes of the same
year. As next step we have classified the pixels of this 20 channel picture
by linear stepwise discriminant analysis (3). In this way we hope to be
able to distinguish each of the given land use categories within the test
site "Grosses Moos".
Misclassifications have to be taken into account by picture-elements of
smaller fields than the resolution and by pixels lying on the border to a
neighbouring field. Since a viewing field is cut by scanlines in a different
manner with each overflight of the satellite quite a large number of mixed
pixels are to be expected.
The following categories were classified (see also Table 1):
lake, reed, urban land, woods, barley, summer wheat, winter wheat, rye,
potato , sugar beet, rape seet, rape, vegetables, pea, meadowland, corn
and vine-yards.
A visual impression of the classified test site is given in Fig. 2 where the
character for each category (code see Table 1, column 2) is printed image-
wise. For easy comparison we have masked those pixels (white areas) where no
ground truth was available.
As a technical detail and example a summary of computational results is
given in Table 2. This was one output of the stepwise linear discriminant
analysis for the 20-channel problem: from this table we can read out the
priorities for the "variables entered" ‘in the routine whilst computing.
Unfortunately there is no physical intuition for understanding this result -
we only know from experience that if we change one of the training sets a
little bit, the hierarchy will change completely.
V" Graund truth digitization
To verify the classification results it is necessary to compare them with
ground truth information. Instead of trying to match visually a satellite-
picture with maps or orthophotographs (e.g. with a zoom-transfer scope), in
order to identify the fields, we have tried a digital method for picking out
samples:
We digitized a ground truth map with the same precision as that of the
Landsat scenes. That means that we have in the end a ground truth pixel
for direct comparison with each classified Landsat pixel. Naturally our
ground truth does not cover the test site completely.
Fig. 3 shows a character-print of the ground truth signatures where each
capital represents a feature corresponding to the codes given in Table 1,
column 2.