The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
552 
Beside genetically guided road key point identification, the idea 
of ellipse clustering was developed to another novel method 
called “increasing ellipse”. In this method, for each road patch 
representative ellipse, the following parameters are determined: 
Area. = 4 .a ; .b ; 
Diff%, = Area -- N i 
Sample% j 
Area i 
N i M 
N 
(3) 
In this equation Diff% shows the percent of the i' h ellipse area 
difference with the area covered by its associated road pixels. 
Also Sample% is a measure of the expected assigned road 
pixels to the i' h road patch.Based on the above introduced 
ellipse parameters, road patches are categorized as Noise, 
Coincide and Under-Evaluation patches.Noise patches are those 
clusters having the following conditions: 
(Diff%) Mean Dtff% % + 2.5<J Dlff% ) A ND(Sample%(0.5) 
Sample%(0.1 (4) 
b( 2 
In equation 5, Mean Diffy an d C7 D ^ % are the mean and standard 
deviation of Diff% i values of all M clusters. 
Coincide patches are those clusters satisfying the conditions 
expressed in equation 5. 
The patch is not markes as Noise patch 
Diff%(Threshold (5) 
a < Ab 
Under-Evaluation patches are the rest of clusters not marked as 
noise nor as coincide patches.Figure 5 shows the methodology 
of the invented increasing ellipse clustering.The termination 
condition of this procedure is to have no ender-evaluation 
patches while all of them are categorized as coincide or noise 
patches 
Figure 5. Increasing ellipse road vectorization 
The termination condition of this procedure is to have no ender- 
evaluation patches while all of them are categorized as coincide 
or noise patches. 
2.2.2 Road Key Point Connection: In order to make correct 
connections between the identified road key points, the 
presence of common road pixels between adjacent road patches 
were used as the connection guide. 
For this reason, a fuzzy shell clustering was implemented on the 
pre-determined ellipse, representative of coincide patches. 
Vague samples, which are road pixels belonging to more than 
one road patch with rather the same membership values, were 
determined based on the obtained membership matrix. Using 
the centroid of vague road pixels as the middle point of key 
points, the corresponding connections were generated. 
3. PRACTICAL RESULTS 
In order to evaluate the functionality of the road extraction 
method proposed in this research, two sub-samples of pan- 
sharpened Quick Birds and IKONOS images from Bushehr 
harbor and Kish Island in Iran were used as case study. Figures 
6 and 7 show the source input images with their manually