theoretical drawback is that Maximum Likelihood assumes
the classes being normally distributed.
One Neural Network method that is a non-parametric clas-
sifier is Learning Vector Quantization (LVQ) by Kohonen
[Hertz, 1991, Kohonen, 1989, Kohonen, 1995]. It looks very
promising because of the efficient training process and its
capability to learn non-normally distributed classes. The pur-
pose of the training process of LVQ is to find a “codebook”
which is a quantization of the training data. This codebook
can be used to classify the entire image by performing a near-
est neighbour labeling process.
5.2 True error rate approximation
Having no extended ground truth the performance criteria for
training is true error rate approximation.
In a first step we divided all training pixels randomly into
reference and testing sets. The sets are given in table 2. 10-
Table 2: Ground truth
forest type | stand age | stand density
all 3145 4598 4631
reference 2661 4088 4142
test 484 510 489
fold cross validation was used to get approximations of the
true error rates of the classifiers. À further statistical mean
of estimating the performance was to carry out one design
and test step to obtain the confusion matrices.
5.3 Verification
Two experts from the Carinthian government and the most
experienced collegue from the institute verified the results vi-
sually. This step was very important because the quality of
the ground truth was not known. Furthermore, our experi-
ence with LVQ was limited at that point of time.
6 MOSAICKING
Due to the spatial distribution of the training areas all over
of Carinthia a seperate training for all satellite scenes was
not possible. For the scenes covering the eastern and west-
ern part of Carinthia respectively, the ground truth was not
covering all classes sufficiently. Thus, first the main scene,
which covers most part of Carinthia, was trained with the
ground truth and classified. In order to establish a classifica-
tion mosaic of all satellite scenes, the classification results of
the main scene within the scene overlay were used as train-
ing areas for the classification of the edge scenes. These had
to be classified separately and combined to a classification
mosaic afterwards.
Table 3: Statistical results
Classification True error rate
Forest type 69.51% + 0.78%
Stand age 65.29% + 0.96%
Stand density | 83.57% + 0.65%
Furthermore, some small areas covered by clouds had to be
replaced by classification results of cloud free scenes. Wher-
ever possible, the edge scenes were used for this purpose.
However, for some parts the central scene had to be brought
604
in using the image not empoloyed in the original classification
of the respective forest parameter.
7 RESULTS
The final approved classifications were done with Maximum
Likelihood for forest type and stand age, and LVQ for stand
density. While the statistical results were better for LVQ in
all three classifications, the experts found problems With the
LVQ results. Small classes tended to be underrepresented, the
overall result was too smooth in appearance. While training
with LVQ problems were encountered with repeatability of
the training results with identical sample sets. Furthermore,
we are suspicious that we did not obtain optimal results with
LVQ. As Song and Lee [Song, 1996] point out in their very
recent paper, mean problems of LVQ are:
1. good initial values for the codebook
2. no garantee for optimal codebook
3. optimal stopping point.
The CV results of the approved classifications are given in ta-
ble 3, the confusion matrices in table 4.to 6. The mosaicking
resulted in complete classified images where the cutting line
remained invisible.
Table 4: Forest type: confusion matrix
(a) | (b) | (c) | (d) | « classified as
77:40: 112 5 | (a): deciduous
31-|321.4. 20 5 | (b): mixed deciduous
8 12 | 56 | 27 | (c): mixed coiferous
7 9 | 12 | 172 | (d): coniferous
Table 5: Stand age: confusion matrix
(a) |l (b) |- (c) | <= classified as
103 | 45 | 18 | (a): young stands
58 | 183 | 36 | (b): mature stands
11 15 | 41 | (c): old stands
Table 6: Stand density: confusion matrix
(a) | (b) | « classified as
54 | 40 | (a): 0 — 6096
43 | 352 | (b): » 6096
8 CONCLUSION
As to chosing the best classifier it is very important to
examine the confusion matrix and - additionally - to verify
the results to obtain the desired outcome. This process was
carried out together with the client which may not be practi-
cal in general, however, this approach was the only possibility
to provide the client with adequate results due to the prob-
lems mentioned.
Neural Network classifiers do have tempting features but also
unexpected drawbacks. We do not recommend to experiment
with new classifiers when there is existing experience, because
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996