Barsi, Arpad
Figure 13. Thematic map produced by considering PCA
and the 4-neighborhood
Figure 14. Thematic classification taking PCA
and 8-neighborhood into consideration
The distribution of the classes is compared in Table 5.
(Abbreviation N is for neighborhood.)
% %
Class 4-N 8-N
F1 19.2 18.6
F2 10.8 8.2
MI 9.8 7.9
M2 28.2 26.2
U 252 33.5
W 6.9 6.2
Table 5. Comparison of the classes for the combination of PCA
and 4- and 8-neighborhood
Two classes (E2 and W) are almost the same as the original, E1
and R1 have slightly more pixels. The most changes are in class
U, which is about the half of the original and the three times
greater R2 class.
The comparison of all designed neural networks can be seen on
a zoomed detail of the map. The detail shows a part of the
original satellite image, where urban (U), meadow (M) and
water (W) classes are common.
a) original image detail
b) classification by the original
bands
c) classification by the PCA
transformed bands
d) classification with
4-neighborhood
e) classification with
8-neighborhood
f) classification with PCA and
4-neighborhood
g) classification with PCA and
8-neighborhood
Figure 15. Comparison of all networks on the same detail
4. CONCLUSION
The experimental project has proved that artificial neural
networks could be designed and successful trained for managing
LANDSAT TM satellite image. Especially the two generally
used image processing tools were studied, namely the principal
component analysis with the connecting image transformation
and the neighborhood.
The chosen project area is covered by forest, meadow, urban
and water classes. In most cases two subtypes were
distinguished. The neural networks require adequate training
and test sets, which were prepared carefully. The training data
were applied for the definition of the right network structure and
the corresponding network parameters. The test set aimed the
quality control of the work.
146 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.
