vation 
wn in 
ulated 
ion of 
ed for 
ine in 
  
Table 1 Test areas 
  
  
  
  
  
  
  
  
  
  
  
Test area name| Land use in 1987 Land use in 1994 
C Bare land Houses 
D Bare land Houses 
E Houses Houses 
G Trees and houses Houses 
H Trees Bare land and houses 
2.2 Results 
2.2.1 Comparison of DEM (1) and (4) — Accuracy test: 
DEM made by automatic stereo matching was compared with 
the DEM measured manually with a stereo-plotter. This result 
shows accuracy measures of automatic matching. The source 
material is old aerial photos. The results are shown in Table 2 
for the test areas. 
Table 2 Comparison of DEM (1) and (4) — Accuracy test 
  
  
  
  
  
  
  
  
  
  
  
  
  
Test |Mean height|Mean height] Difference | Standard 
area of (1) of (4) of the mean | deviation 
(automatic) (manual) 
1 H 2 d ag 
C 78.66 78.49 +0.17 1.00 
D 87.47 87.57 -0.10 0.92 
E 78.57 78.81 -0.24 2.06 
G 31.35 31.68 -0.33 2.35 
H 43.84 46.55 -2.71 4.49 
unit: meter 
Here H, and Hz are the mean of height inside the test areas. 
These statistical parameters are defined as 
n=61x61 
H,=Sh,/n 
Sie 
i=in 
o- [Sh -h,, - dy! /(n-1) 
where h, ; and h2,; are height at grid point i. 
  
It can be seen from the result that the differences of the mean 
height value are small and standard deviations which are the 
indication of accuracy of automatically derived height are 
within about two meters except for tree area. Therefore we 
consider the DEM of automatic stereo matching has enough 
accuracy for the purpose of-change detection except for tree 
areas. 
2.2.2 Comparison of DEM (1) and (2) — Change detection 
by the difference of DEM from stereo matching: 
DEMs at two different time obtained from stereo matching were 
compared against each other in order to test the feasibility of the 
method. The results are shown in Table 3. 
Table 3 Comparison of DEM from stereo matching 
  
  
  
  
  
  
[ Test [Mean height| Mean height [Difference] Root mean 
area of (2) of (1) of the | square error 
Hi H, mean r 
Leur |. (1994) (1987) d 
C 79.89 78.66 +1.23 2.23 
D 88.72 87.47 +1.25 2.05 
E 79.05 78.58 +0.47 1.90 
G 33.42 31.35 +2.07 3.49 
H 40.38 43.84 -3.46 6.40 
  
  
  
  
  
  
  
549 
  
IS 4A (hy; ml) /(n-1) 
Root mean square error r is defined slightly different from the 
above "standard deviation" 0. This is because the difference 
itself has the meaning that height change occurred during the 
time interval in this case but in the former case the height 
should be the same and the difference is considered as an error. 
It can be seen from the Table 3 that test areas C and D where 
bare lands turned to housing areas have average height increase 
due to newly constructed houses whereas change of average 
height is small in area E (40.47 m) where no land use change 
occurred. Area G changed from trees and houses to housing 
area. Cutting trees decreases height and house construction 
increases height. It seems mixed effect of these appeared in area 
G. Trees are cut widely in area H and land is prepared for 
building houses. This situation results in different values in the 
mean height difference and the root mean square errors. 
Therefore it can be said that change of average height or root 
mean square error in an area between two time may be used as 
the indication of land use change when height is measured by 
automatic matching of stereo pair photos. But there remains two 
mutually related problems. One is to determine the threshold 
value to judge change has occurred in the area. The other is to 
determine an appropriate area size. If the area where statistical 
parameters such as a mean and a root mean square error are 
calculated is large, then many kinds of land use and land use 
change patterns are included in one area and this method does 
not work well. On the other hand, if the area is small, errors in 
height measurement cause much erroneous detection of change. 
If the result of stereo matching become much reliable, then it 
will be possible to use height data at every grid point to detect 
changes. 
2.2.3 Effect of area size: As mentioned above, it is desirable 
for this method of change detection that land use is as uniform 
as possible in a selected test area. When the area is smaller, the 
land use in the area is more uniform. Therefore the effect of area 
size is examined by dividing the test areas into four and nine. 
The difference of mean heights between two periods are shown 
below for the original and divided areas. 
Table 4 Difference of the mean heights 
From left to right: original area (150 m square), four division (75 m 
square), nine division (50 m square). 
Area C (Bare land — Houses) 
1.96 0.89 
  
2.38 | 1.26 | 0.18 
2.12 | 0.98 | 0.79 
1.20 0.87 1.77 11.231 0.56 
  
1.23 
  
  
  
  
Area D (Bare land Houses) 
1.17 1.18 1.63 | 1.46 | 0.56 
0.69 | 1.18 | 1.51 
1.32 1.37 1.61 | 1.07 | 1.61 
  
  
1.25 
  
  
  
  
Area E (Houses — Houses) 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
0.62 0.46 0.46 | 0.48 | 0.31 
0.47 1.41 | 0.28 | 0.65 
0.73 0.17 0.80 | 0.09 |-0.04 
(continued) 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996