ts (settle-
riangles to
al
ialyse
en original
acted seg-
elements a statistic can be created which can be used later in
a decision process to model uncertainty. We achieve a better
representation of the object shape (fig. 6) and automatically
a raster/vector conversion will be achieved.
Example 'Forest': |n a similar way segments of forest areas
can be created. The DLM200-objects enable to determine
the spectral reflectance values of all forest areas in the im-
age. Assuming that a majority of the captured pixels are valid
a histogramm analysis allows to extract the main reflectance
behavior, where disturbances, e.g. by stub areas or digitizing
errors, are excluded. The related pixels get marked in the
same way as ‘settlement’. The resulting Delaunay triangu-
lation is shown in fig. 7. After selection and fusion of valid
triangles we get the segments for forest areas (fig. 8).
Image: tm83.c[ 1 1 100 100] ê
Figure 7: Segmentation of forest areas by Delaunay
triangulation (subset of test area 'Speyer')
Image: tm83.c[ 1 1 100 100)
Figure 8: Selection and fusion of valid triangles to
segments (forest, see fig. 7)
The other object classes (e.g. water) can be treated in ana-
logous way.
755
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
3.2 Texture Feature
As mentioned in chapter 2 its necessary to expand the fea-
ture base in satellite image analysis, because using only the
feature spectral signature doesn't deliver satisfying results.
The first extention may be texture as another spectral — but
object-oriented — feature. The investigation of different tex-
ture parameters was started with the most common and fa-
mous Haralick parameters (homogeneity, mean, entropy, con-
trast). Out of these, acceptable (but not optimal) results can
be obtained by homogenity. Fig. 9 shows the parameter val-
ues for different object classes and different training areas.
E Homogeneity legend
250} 4 - settlement
U - agriculture
+ + O - forest
200+ + + + - water
+ + *
+ +
+
r T *
& 150 re *
2 + +
"S + ++ o
a 49-0 OF o «o
100, + S m o 9
+
c a % Vo 009 500 ©
a a © oo,
sol a a go 00 o $9 o
at e
n. ^5Ba^4 ^
^
E zi i
Ü 50 100 150 200 250
Band 1
Figure 9: Haralick parameter homogeneity as texture
parameter for different object classes and different
training areas
3)
S
2l
$8 :
&
26]
A^
| ^
22 -] A
| A
A
18 -] A
Oo
14
D
= o
a O
10 -] a Oo
ü 0
1 Bu A - settlement
6 7| au Oo - forest
7 t D - agriculture
25] =
+ - water Banda e.
LT o Ll ul T reyvalue] -
2 6 10 14 18 27 € Tgreyvalue]
Figure 10: Modified standard deviation as texture
parameter for different object classes and different
training areas
Especially the poor separation between settlement and agri-
