'stanbul 2004
ca
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
€ Forest
YF older than 7 years
AA tree
EE on ® OF coniferous up to 45 years
Se) -@ OF coniferous older than 45 y
a qu OF deciduous
N p. UA tree
€ ) Non-forest
: UA green areas
E AA fields
YF forest up to 7 years
OF forest up to 7 years
10 Urban area
| Ea UA road
i UA dark house
rt UA light house
UA house
AA house
AA road
YF road
Figure 3. Lower-level classification
Results shows that certain parts of roads were classified into
class houses. Pixels from class houses were classified into trees,
roads, and fields. Classified class called tree comprised pixels
from houses, roads, and fields, class roads was a class having
also pixels from houses. Forest classes showed very good
results for the coniferous forest. The forest younger than 7 years
old was classified partly into the deciduous forest. The best
results were obtained for fields. Results reliability is higher
between urban areas and forests or urban areas and fields.
Smaller reliability was found between urban classes — houses
and roads, etc.
4. CONCLUSION
The B&W photograph classification can be performed with a
relatively high accuracy. There are three necessary conditions
offering such good results. New channels have to be calculated
from original data (photograph) — channels calculated by
filtering where median and Gauss filters are used, and channels
using Haralick functions. Careful testing of kernel sizes for
filtering and careful choice of filter window sizes for Haralick
functions compared to final classified object sizes can
sufficiently improve classification accuracy. The presented
project used kernel sizes smaller than the smallest classified
objects. Filter window sizes were smaller, equal and a little bit
larger than these objects. The higher number of Haralick
functions improves classification accuracy.
Object-oriented analysis using image segmentation followed by
segment classification is the second necessary condition for
good results of classifications.
Two level segmentation in the reverse order of segmentation
(from higher to lower) is the third condition. The higher-level
segmentation ensures fragmentation of image data with
overlapping pixel values for different classes into thematically
closer and smaller image parts whose further segmentation and
classification offers good results. The lower level classification
is performed individually for already fragmented parts of image
data.
'The confusion in resulting classification can be found in urban
areas. The overall accuracy was about 90 per cent. However, the
accuracy for individual classes varied from 50 per cent to 100
per cent. Low accuracy values were in case of two classes —
trees whose pixels were classified into deciduous and
coniferous forests. The class tree did not distinguish tree types.
It was a class describing individual trees or tree groups out of
forest areas. Rather wrong distinguishing can be found between
very young forest and deciduous forest. It is difficult to
distinguish these two classes even during visual interpretation
especially for forest age younger than 40 years.
REFERENCES
Baatz, M. and Schápe, A., 1999. Object —oriented and multi-
scale image analysis in semantic networks. Proc. Of the a
International symposium on operationalization of remote
sensing, August 16-20, , Enschede ITC.
Borisov, A.N., Kashin, V.B., Khlebopros, R.G., 1989. Method
for indication of horizontal structure of tree stands. Doklady —
Biological Sciences, 1989. 293: (1 — 6), 132 - 133.
Halounová, L., 2003. Textural classification of B&W aerial
photos for the forest classification. Proc. of the 23". symp. of
EARSeL, Gent, Belgium, June 2-5, 2003: 173-179.