ISPRS, Vol.34, Part 2W2, “Dynamic and Multi-Dimensional GIS", Bangkok, May 23-25, 2001 
106 
The results were extremely accurate; the registration mismatch 
was just a few pixels on average. 
Figure 7 The reference pixels in the satellite image 
25000 
15000 
-50 -20 10 40 
Displacements (Pixel) 
Figure 8 The results of voting scores 
TABLE 1 The mean square errors yielded by proposed method 
Region 
1 
2 
3 
4 
5 
MSE 
2.17 
1.91 
2.23 
1.57 
1.48 
Feature Pixels 
181,582 
80,864 
92,069 
65,451 
48,343 
Figure 9 The results of displaced map outlines 
We also tested the approach of using image edges as matching 
features to quantify the advantage of the NDVI approach. The 
edges were extracted using the Robinson operator. Fig. 10 
shows the edges extracted from the satellite image in Fig. 1. 
Fig. 11 shows the voting scores using these edges. No clear 
peak was detected in this case. Additionally, the mean square 
errors are calculated using resulting displacements of our 
approach. Table 2 shows the results of the 5 regions. The 
number of feature pixels was about 910,500 per region on 
average, and the average error was about 19.53 pixels 2 . There 
are a large number of feature pixels, which include many pixels 
unsuitable for voting. Therefore, the error is about ten times 
worse. This fails to well determine mismatch. This is also the 
reason for the large computation time. These results reveal that 
the NDVI approach yields higher accuracy and smaller mean 
square error than the approach of using image edges. It is clear 
that our approach can extract effective features from satellite 
images because it makes good use of topographical reference 
objects in voting. 
Figure 10 The results of edge extraction 
-20 10 
Displacements (Pixel) 
Figure 11 The results of voting scores using edge extraction 
TABLE 2 The mean square errors yielded by edge analysis 
Region 
1 
2 
3 
4 
5 
MSE 
16.26 
16.32 
29.49 
16.84 
18.77 
Feature Pixels 
1,360,398 
692,913 
1,218,203 
668,366 
612.646