+c, eg, =" line(8 , ,  , )' ^ 
secorridor A0, 20; ^K, -Kj 
C4 (6,)7 cj  ife, ="line(®,,x) otherwise (3.4) 
C 
S Ll" ani 
a if €, ="no-line 
c. is an "additional current" due to the presence of a GIS 
object. The membership of a site in a corridor and the state 
which is affected by this are computed prior to the line extrac- 
tion and stored in a raster map. 
4. LOCAL DATA EVALUATION 
According to (2.10) we are dealing with two types of data: the 
SAR intensity and the interferometric coherence. Both are 
evaluated in a similar manner which is treated in this section. 
For edge detection in SAR intensity data the ratio edge detector 
has been shown to give the best measure of edge strength. It 
corresponds with the multiplicative noise characteristics of the 
data and results in a constant false alarm rate (Touzi et al., 
1988; Bovik, 1988; Adair & Guindon, 1990; Caves, 1993). Our 
approach is also based on the ratio detector. 
In general the ratio operator compares two small regions of the 
image, e.g. the left and the right half of a window. In each of 
the two regions the averages of the intensities are computed. 
The normalized ratio output is the ratio of the two averages 
using the larger one as the denominator: 
minui) 
max(( )2)) 
where (A) and (45) are the mean intensities. If r is close to 1, 
(4.1) 
the regions do not contrast. If r approaches 0 and the intensity 
is homogeneous in both regions, there is a discontinuity along 
the boundary of the regions. 
To detect a line a detector mask is defined which consists of 
three regions (Lopes et al., 1993): a line region and two regions 
at the sides of the line region (see Fig. 4.1). Curved lines of 
varying widths can be considered. The detector masks are 
generated prior to the processing of the scene for a set of line 
directions, curvatures and widths. They correspond to the 
discrete line directions and curvatures defined by (3.1). 
The ratio operator is applied to all combinations of the detector 
mask regions. The line detector has to be prevented from 
giving a response to the location of edges or strong scatterers. 
Therefore, we follow a procedure similar to the one proposed 
by Lopes et al. (1993). In case a line is present at the processed 
location the line region is homogenous, which means that the 
pixel intensities do not vary much. For SAR intensity data, this 
condition can be checked by computing the coefficient of 
variation 
nog (4.2) 
(7) 
316 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
  
  
  
  
  
Fig. 4.1. Geometries of detector masks for line detection. The spot in the 
center of the window marks the site to be investigated. 
where (T) is the mean intensity and 6, is the standard devia- 
tion of the intensity I. As &;=I for homogenous regions, v; of 
the investigated region has to be less than a threshold some- 
what larger than 1 if the region does not contain any structures. 
As the average values of the side regions may be influenced by 
the presence of strong scatterers, they are replaced by the 
median of the intensity values. It is checked whether the line 
intensity differs significantly from the intensities of both side 
regions by computing the ratios between the line region and the 
side regions. These ratios are tested for membership in the 
ratio distribution of regions without contrast, i.e. regions which 
have the same: intensities. For this purpose we adopted a 
threshold derived from the PDF of the normalized ratio r which 
is based on SAR intensity statistics (Caves, 1993): 
N,L 
ey 
T(N;L* NL) [s m 
1,,1,,N,N3,L)= 
pli; 23V 31V) ) ey T ved 
1+—. 
  
Ij Ny 
bas 
LN, 1 
* N,L+N,L | (4.3) 
1411 Na 
ly. N, 
I; and I, are the intensities, N; and Nj are the number of 
pixels in both regions, and L is the number of looks per pixel. 
L is computed by dividing the ground pixel size by the size of 
the resolution element of the sensor. Thus N;L is the number of 
independent samples. Fig. 4.2 shows the PDF for various 
contrasts I;/I. It is strictly valid only for ratios of averages and 
not of medians, but tests have shown that it is a good approxi- 
mation. Two regions are considered to be significantly different 
if rer L where r 1 is a threshold derived from 
joe = 1,,N;,N,,L)dr = 0.05. (4.4) 
0 
If the line region is significantly different from both side 
regions, a line site has been found. Then the side regions are 
checked for similarity which is assumed, if r>r, where r2 is a 
threshold derived from 
S ol = 1,,N;,N,,L)dr=0.5. 4.5) 
0 
    
  
    
   
  
  
   
   
  
   
  
  
  
   
   
  
   
   
   
   
   
   
  
   
     
    
  
   
  
   
   
    
   
   
  
  
  
   
   
   
  
  
   
   
   
    
If the s 
operato 
line reg 
the les: 
commo) 
the unit 
From t 
intensit 
of a pb 
where ı 
A reas 
assume 
H is co 
where 
normal 
normal 
poses « 
ton al 
more, 
geneit 
strictly 
(4.7) v 
Cp, is 
scene. 
For no 
distrib 
This ic 
of mul 
distrib 
densit 
multis 
corres 
T can 
line pi 
image