The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
525 
SI Class Confusion Matrix 
File 
Confusion Matrix: [Memory5] (400x400x1) 
Overall Accuracy = (141450/160000) 88.4063% 
jKappa Coefficient = 0.8440 
Figure 3 Precise Analysis 
2.3 The research of road extraction experiment 
2.3.1 Morphology processing: Carried on morphology 
processing using the mathematics morphology operation, such 
as Open, close, erode, dilate and so on. First change the 
clustering image into binary image using the morphology 
function- im2bw, then process the binary image with the 
mathematic morphology algorithm. We can see from the binary 
image(see Figure 4), Around the path there is much disturbance 
which is caused by the spectrum error such as " different objects 
with same spectrum ", "same object with different spectrum" 
And the roof that has the same spectrum as the road and so on, 
so in order to extracting the road net information to reject 
these disturbances is very important. After repeatedly 
experiments we select the long line structure element according 
to the road structure, seO=strel('line', 12,30), 
sel=strel('line',10,120).Separately carried on the close operation 
using the two structure elements , remove the tiny noise, at the 
same time ,separate noise that adheres to road information , this 
step is especially important, in a sense, the selection structural 
element has decided the road information in extraction scale. In 
this image the road information is quite tiny in disorder, 
therefore we use smaller size structural element, maintain detail, 
of the road .But because of the road information geometry 
characteristic complex and changeable, so the structural element 
used in this paper cannot use in common, we can select 
appropriate structural element according to the road collective 
situation, or choice different structural element to carry on 
processing many times, in order to achieve the most superior 
effect. After the close operation, the main road information 
displays separately in the close operation result images, but 
there is also the non- road information outside the branch road , 
which is caused by the segmentation result ,we may remove the 
irrelevant information using "bwareaopen "function, in order to 
obtain the integrity road information, after we add the two close 
operation result images together to obtain the integrity road 
information, the we use thin function to obtain the road median 
line, the image of figure 5 is the road median line image. 
2.3.2 Using the seed growth algorithm to eliminate the 
short line, extract the road median line that has certain 
length and direction: We can see from figure 5, we have 
already roughly obtain the road median line, but because ofthe 
image itself characteristic and mathematics morphology 
Limitation, there are many irrelevant short lines, in order to 
eliminate this irrelevant short lines, extract more precise road 
information, the seed growth algorithm is used in this paper. 
From the image we can see the image DN is 0 and l,we put the 
image into a two -dimension matrix, the road DN is l,we search 
the road information according to the following rules: 
Figure 5 Road median line extraction 
First, looks for a pixel point whose value is one, judges whether 
it had been searched, if not, then it is as seed growing point; 
Second, Store the seed points to the matrix -road seed (),then 
search seed point's eight neighborhoods to look for the pixel 
elements whose value is one, if this point has not searched then 
it is put into another matrix seed (); 
Third, take a point from the matrix seed() as a seed point, 
duplicates the second step, until the matrix seedQ is empty; 
Forth, Judge the matrix road seed(),if the element integer is 
bigger than five pixels elements then it belongs to the road 
median line, otherwise rounded down; 
Fifth, store the new matrix roodseed() into the another matrix - 
rood(), and clear the matrix of roadseedQempty, then redundant 
one step , until all pixel elements of the matrix road() have been 
scanned; 
Sixth, read-out the matrix road(), then output a binary image, 
this image is the extraction road median line. 
The image of figure 6 presents the extraction result. Figure 7 
shows the superposition result, the Figure 8 is another 
experiment area.