detection of such points the so-called Geometrical Ratio 
Operator (GRO) has been implemented. In this context 
the work carried out by Desnos and Matteini (1993) or 
Touzi et al. (1988) has been considered. 
2.1. Prefilterin 
In a first preprocessing step prefiltering algorithms may 
be applied to the images to be treated in order to reduce 
the SAR inherent speckle noise. For that, commonly used 
adaptive filter methods may be used like, for instance, the 
Frost, Lee, Kuan, or MAP filter, all of them being based 
on local statistics. These filters reduce the speckle noise 
as a function of heterogeneity measured by the local 
coefficient of variation. The MAP filter additionally 
considers geometric informations. 
2.2. Preclassification 
In another preprocessing step the image information may 
be classified into homogeneous areas, where the SAR 
image signal shows little variation, and into 
heterogeneous areas where the SAR signals varies 
significantly due to textured areas, edges or point targets. 
Subsequently, homogeneous areas can be considered to 
be of no further use in this concern and can be excluded 
from further processing. This may significantly reduce the 
computational effort for the subsequent procedures. 
A very efficient and robust index for textural information 
which measures the image heterogeneity is given by the 
local image coefficient of variation (COV). Image 
reduction methods based on the Gaussian kernel filter 
may be applied in order to eliminate micro structures in 
the SAR scene and to make the detection procedure 
more sensitive to macro structures. 
A frequent problem particularly in built-up areas is the 
presence of strong scatterers, being small bright features 
with point-like shape. These in general could represent 
well defined tie-points, but are very critical within the 
subsequent tie-point matching procedure as their shape 
may be different in other SAR images or they even may 
completely disappear. A possibility to identify and 
eliminate strong scatterer areas was therefore included 
into the preprocessing and preclassification procedure. A 
selected percentage of the brightest pixels in the image is 
therefore interpreted as strong scatterers and region 
growing is applied in order to mask respective areas. 
2.3. Geometrical Ratio Operator 
  
The SAR speckle noise makes the extraction of textural 
features very difficult. As gradient operators are strongly 
dependent on the main reflectivity, ratio operators are 
considered to be more appropriate. Therefore, the so- 
called Geometrical Ratio Operator (GRO) has been 
developed for the detection of tie-point candidates. A 
particular objective of this operator is the detection of 
structural information like edges between extended areas 
and curvilinear structures like roads, shapes, hedges, 
316 
shadows etc. Further advantages of this operator are the 
constant alarm rate, the possibility to implicitly determine 
threshold parameters and its simple implementation. 
The ratio operator is defined as the ratio of the pixel 
average of two non-overlapping neighbourhoods on 
opposite sides of the inspected point. To be sensitive to 
any geometric location and orientation of edges, the filter 
is applied in the usual four directions, i.e. horizontal, 
vertical, and twice diagonal. Using neighbourhoods of 
increasing size in combination with variable thresholds 
allows to detect both micro edges as well as extended 
features. 
Subsequently, the four directional GRO results are 
exploited in order to detect corner points. Such points are 
defined by their high probability to be an edge in at least 
two directions. Finally, the most useful tie-point 
candidates are selected taking into consideration a given 
minimum distance between two neighbouring points. 
2.4. Tie-point Detection Example 
The principle of the tie-point candidate detection 
procedure is demonstrated for a small SAR image chip of 
200 by 200 image pixels. This image chip shows a major 
river body appearing in rather dark grey values (see 
Figure 1). Figure 2 shows some examples of intermediate 
results of the various optional preprocessing steps 
mentioned above, like particularly: 
e Kuan-filtered SAR image with a 3 x 3 filter kernel. 
e Preclassified image chip showing the separation of 
the image into homogeneous areas (black) and 
heterogeneous areas. 
e Removal of areas containing strong scatterers, which 
were defined in this example by the 0.01 brightest 
pixels in the image chip. Region growing with 7 pixels 
was further applied to define respective strong 
scatterer areas. 
  
  
  
  
Figure 1: Selected image chip. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996