indaries 
1, 
hand 
sin a 
yf the 
stiga- 
ral 
ctral 
simple 
class 
and 
ts 
posed 
don 
ements. 
ments 
ck 
of 
fferent 
ted 
mina- 
cal 
using 
fig.13. 
cen 
  
E 
  
ORIGINAL IMAGE , 
| 
preprocessing 
data reduction 
| (1) 
automatic training 
| (3) 
multispectral 
classification 
  
  
  
  
(2) 
  
  
  
  
(4) statistical 
textur analysis 
in raster elements 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
(5) (6) [ (7) (8) 
multispectral 
clustering 
(9) 
structural (10) 
textur analysis 
  
  
  
statistical 
textur analysis 
in segments 
(11) (12) 
  
  
  
  
  
  
  
| (13) 
complex large 
TEXTURAL OBJECTS 
large small 
SPECTRAL OBJECTS 
PRIMITIV 
TEXTUR 
Fig. 13: Evaluation system 
After preprocessing and data reduction (1), an automatic training 
wil 
a) 
wt 
uu SE er rt RS 
l be executed. "hree cases can be distinguished* 
If the image is mainly composed by spectral homogeneous featu- 
res, the large and strong homogeneous regions are selected (3) 
as training areas for the following multispectral classifies- 
tion. Large and compact spectral objects are extracted (5). 
Smaller objects (6) may ‘be primitiv element of a texture region 
and a structural texture analysis will verify this hypothesis. 
If not the whole spectral features are yielding in °the training 
set a reject class occurs (4). Here a multispectral clustering , 
generates the spectral homogeneous elements (9) for the struc- 
tural texture analysis. 
If the image is mainly composed by textured regions no training 
areas result (2). Here the statistical texture analysis in 
raster elements obtain (7) a reliable presegmentation of an 
image. In the different segments of homogeneous texture 
features the multispectral clustering is simplified because 
of specific texture. The following (9) structural analysis 
search for similar properties of the texture elements. Line 
shaped and point like objects are detected (11) as well as 
primitiv textured regions (12) yielding only one texture 
element. 
41 
rr