" A DT SEE 
ing areas for. the multispectral analysis will be detected 
automatically, without a priori knowledge. Here the complementary 
properties of a contrast extraction algorithm and a region growing 
algorithm are exploited... 
For the texture analysis we can distinguish between the 
statistical and the structural texture analysis. The main 
properties of these methods are.discussed here. Investigations 
about the statistical texture analysis show that an essential 
problem of the statistical texture analysis is the dependence 
of the extracted features from the size and shape of the defined 
raster elements. It will be demonstrated that the shortcomming 
of the statistical texture analysis methods can be decreased by 
an additional combination with the structural texture analysis. 
Based on the results of the investigations a complex processing 
system exploiting the different types of properties and combining 
the complementary methods will be proposed. ; 
COMPARISON OF MULTISPECTRAL AND TEXTURAL METHODS 
With regard to the combination of the multispectral and textur 
analysis, their principal properties will be compared first. 
Investigations /1, 2, 5, 6/ demonstrate the characteristic pro- 
perties of the different methods: 
- a) multispectral analysis: 
- Complete classification with high resolution of spectral 
homogeneous objects 
- Additional classification of line shaped objects and point 
like objects 
- Bad classification of textured areas with a large part 
of reject 
- b) statistical texture analysis: 
- À better detection of objects with essentially textural 
features 
- A simultaneous detection of spectral homogeneous objects 
- A gmall resolution on object boundaries 
- No detection of line shaped and point like objects. 
- c) structural texture analysis: 
- Detection of line shaped and point like objects too 
- High resolution of detecting region boundaries 
- Difficult detection of spatial relationships with dominant 
statistical nature 
- High expense 
The aim is a cumulation of the advantages of the different 
methods. This may be realized by a spatial selective use of the 
different methods dependent upon the kind of objects to be pro- 
cessed. This assumes the knowledge about the spatial probability 
distribution of objects in the image. Without a priori knowledge, 
this information have to be extracted step by step with a suitable 
sequence of different segmentation methods. 
36 
The pro 
to dist 
is an a 
homogen 
a contr 
algorit 
of regi 
areas o 
image, 
methods 
Normaii 
are not 
up to t 
some la 
are sel 
are sim 
for 8 m 
the prc 
boundar 
were nc 
To demc 
aerial 
Ti de € 
the res 
high (t 
trainir 
The rec 
documer 
class i 
well as 
  
BALA 
white. 
po M.