The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3h. Beijing 2008 
<«> (0 
3.3 FCM Clustering in Spectral Domain 
In this step of proposed algorithm, FCM clustering technique is 
implemented by defining number of cluster centers (C) and 
fuzzy parameter m in spectral space, in this work, cluster 
centers were not selected randomly. For definition of cluster 
centers, at first C-Means clustering technique was done by 
defining C and the output of this technique were entered the 
FCM as input cluster centers. The number of cluster centers (C) 
is defined according to diversity of image objects and by try 
and error operation. 
After performing the clustering technique, by supervision of 
human expert and definition of suitable threshold, road class 
binary image is obtained. 
3.4 Post-processing of Images 
After selection of road class binary image, since the obtained 
class is not clean and misclassified pixels are in the image, post 
processing operation is mandatory in order to increase 
classification accuracy. 
In this research morphological operators are used to improve 
the image by removing noise and filling gaps. These operators 
consist of Opening, Closing, Bridge, Clean, Majority and 
Thicken which are used according to structure of misclassified 
pixels in the road class binary image. 
3.5 C-Means Clustering in Spatial Domain 
According to section 3.3, the FCM clustering was done on the 
grey level values of the pixels in spectral bands of the image. In 
this step, C-Means clustering is performed on the spatial 
coordinates of the pixels. So initially, a uniform grid of points is 
overlaid on the road class binary image and by using C-Means 
clustering; the location of some predefined nodes is changed. 
These points are called "Active Nodes". Nodes that their 
locations remain unchanged are called "Dead Nodes". Dead 
nodes are simply disregarded for the remainder of the 
process. 
Grid spacing is based on a pre-specified (e.g., operator 
provided) and differs in different images based on the shape 
and width of roads in the image. 
3.6 Using MST to Extract Road Centerline 
In order to extract road centerline, Active Nodes should be 
connected using MST algorithm. In this application Kruskal 
algorithm is used to connect Active Nodes and defining road 
shape. Euclidean distances of all Active Nodes from each other 
are computed and introduce to the Kruskal algorithm as the 
weights. Based on what was mentioned in section 2.3, the 
minimum spanning tree which connects all the Road centerline 
nodes, are obtained. 
This product can be imported into GIS environment in vector 
format. 
3.7 Accuracy Assessment 
Accuracy assessment of the proposed road extraction system is 
performed in two steps: Classification accuracy assessment and 
Road centreline extraction accuracy assessment. 
Classification Accuracy Assessment 
In this step, Confusion matrix is made using Ground truth data 
and classified data and accuracy assessment parameters such as 
Overall Accuracy (OA), Kappa coefficient, User Accuracy 
(UA), Producer Accuracy (PA), Commission & Omission Error 
O C 
( e , e ) are calculated. 
These parameters are calculated before and after post 
processing of images by means of morphological operators. 
Road centreline extraction accuracy assessment 
In order to obtaining the differences between manual and 
automatic extracted road centerlines, distances between these 
lines are measured in location of Active Nodes. Then road 
centerline RMS Error is calculated using these measurements in 
x and y directions. 
4. EXPERIMENT WITH IMAGERY 
After getting familiar with the proposed semi-automatic road 
extraction system procedure, the time is ripe to explore the 
practical results of implementing the system on test images. 
Fig. 3 denotes the procedures of semi-automatic road centreline 
extraction from pan-sharpened IKONOS image of Lavasan city 
in Iran (lm resolution and 119x170 pixels) in mountainous 
area. Fig. 3b shows the FCM clustering of input image and 
resulting road class by defining threshold. It is clear that 
the details of the clustering are immaterial for the 
purposes of this investigation but required parameters are 
presented in the table 1. Fig. 3c shows the result of using 
morphological operators as post processing operation. 
What is important is the existence of classification 
accuracy assessment parameters before and after post 
processing which are accessible in table2.