ISPRS, Vol.34, Part 2W2, “Dynamic and Multi-Dimensional GIS”, Bangkok, May 23-25, 2001 
105 
made objects possible. Note that there is a limit as a size of 
object can be discriminated, which depends on a resolution of 
the satellite image. 
Figure 2 An example of 1/2,500 scale map 
Figure 3 An example of red channel image 
Figure 4 An example of near infrared channel image 
Figure 5 The results of NDVI calculation 
Figure 6 The extracted feature pixels 
Following the procedures described in section 3, mismatch was 
determined using the resulting feature pixels To simplify the 
discussion, only the results of y-direction displacement are 
shown below. Fig. 7 shows the resulting reference pixels in the 
satellite image; this was realized by projecting map objects onto 
the image using known coordinates. This figure shows that the 
projected positions are somewhat shifted against the real object 
positions shown in the satellite image. Fig. 8 shows that the 
voting scores exhibit one clear peak. 64,251 votes were cast in 
this case. The displacement with the lowest mean square error 
is -13 pixels in this case. This value is equivalent to about 52 
meter mismatch. The mean square error is also 1.57 pixels 2 . 
Similarly, the registration results are compared in Fig. 9 to the 
original projection shown in Fig. 7. The figure shows that the 
registration of this image has slight errors. The results show 
that our approach is very effective in suppressing many of the 
errors common in mismatch determination. 
To quantify the average performance of the proposed approach, 
the results of 5 regions (including the above area) are shown in 
Table 1. The size of each region is approximately 500 meters x 
500 meters. In this case, the number of feature pixels was 
about 93,600 per region on average. The average error was 
about 1.87 pixels 2 . We confirmed by projecting the maps over 
the satellite images that the registration has only slight error.