Area G (Trees and Houses — Houses)
0.69 3.47 0.15 | 0.07 | 3.99
2.36 | 2.20 | 4.07
2:37 1.83 2.64 | 1.48 | 1.84
2.07
Area H (Trees — Bare land and houses)
-2.71 | -4.93 -2.99|-4.21|-1.28
-1.29|-6.54|-8.45
-1.18 | -5.17 -0.58|-4.21|-1.94
-3.46
It can be seen from these results that in area C and D, difference
values have positive value and vary larger among divided areas
when the division become smaller. This can be attributed to the
fact that the ratio of newly constructed houses is different
within the small divided areas. On the other hand, area E has no
land use change in the period but shows mean height changes of
40.17 to 40.73 m in four division and -0.04 to 41.41 m in nine
division. These errors are considered to be some mis-matching
in stereo matching and are compensated by taking the average.
From this example, we temporarily drew a conclusion that
appropriate area size is 75 m square and threshold of mean
height is about 0.8 m for finding changes from bare land to
houses.
2.2.4 Comparison of DEM (1) and (3) — Trial to recognize
land use by the difference of height data:
By comparing DEMs of the old photo and of the map, houses
and school buildings are visualized because height drawn on a
map is ground height whereas height obtained by matching is
the height on top of buildings. But it was difficult to get
meaningful criteria to recognize land use from such statistical
approach done in as sections 2.2.1 and 2.2.2. This comparison,
however, should be useful to visualize large buildings and
houses when normal ground height DEM is readily available. It
may also suggest the extent of errors of DEM obtained by stereo
matching.
2.3 Discussion
The comparison of DEMs at different time derived by automatic
stereo matching gives information of land use change. It was
found that about 75 meters square is appropriate for the area
size of taking average for change detection from bare land to
houses and the threshold is around 0.8 meter. It is still to be
done to find out appropriate area sizes and threshold criteria of
mean height change for various kinds of land use change
patterns.
If height value at each grid point is much more accurate, we can
proceed to check the change at each grid point instead of taking
average of large area. Therefore it is an important issue for this
research to examine the accuracy of automatic stereo matching.
We consider the following topics should be studied in this
regard: (1) to carry out stereo matching using other digital
photogrammetric instruments or matching softwares and
examine the accuracy of obtained height value, and (2) to check
the effect of scanning pitch and scale of aerial photos; the
height accuracy seems to be improved when sampling pitch on
the ground is smaller, namely by using large scale aerial photos
and minute scanning pitch.
Another issue is to combine every available data source with the
height data from stereo matching. For example, if height data
are combined with digital cartographic data with boundary line
of buildings, then new construction and disappearance of
houses might be detected one by one. These issues should be
the research items in the next phase of the study.
3. CLASSIFICATION USING COLOR INFORMATION
The objective of this method is to utilize color information of
color aerial photos whenever they are available, for reducing
human work load in change detection. Change of roads could be
visualized after several steps of image processing, including
color reduction, applied to color aerial photo images.
3.1 Background
The basic idea was to use color aerial photos and to enhance the
image by restricting number of colors into very small limited
number such as 8, 16, 24, or 32. It was found that this color
reduction can help recognition of change by human operator to
some extent. It was also found that cubic close-pack algorithm
was better than popularity algorithm or median cut algorithm as
color reduction algorithm. But automatic change detection is
still difficult because color or gradation of aerial photos varies
by various photographing conditions such as season, time of a
day and so on, and there is no easy way to establish
correspondence between colors of photos at different time.
Based on this experience, we decided to restrict our target here
to change detection of roads, important objects in a map. Roads
are considered to have less color variation on image than any
other objects. The method is to give knowledge by human
operator to the result of color reduction.
3.2 Method
Change detection of roads from color reduction was done as
follows. Color aerial photos taken in 1979, 1984 and 1992 on a
scale of 1:10,000 were scanned at 400 dpi resolution. Test site
was Kodaira city in Tokyo. Number of colors was reduced to 24
colors by cubic close-pack algorithm. Then the colors
corresponding to roads were selected manually for image of
each year. After eliminating salt-and-pepper noise by
expanding-shrinking and smoothing processing, elimination of
components other than roads, for example parking areas, was
executed. This was done as follows. At first labeling was carried
out according to color code of each pixel. Then label numbers
of 4-neighbor of a pixel are examined and the number of pixels
having the same label number was summed up for every pixel in
a label. The sum was divided by the total number of pixels
within the label. If this ratio exceeded some threshold value
(value used were 6.25 - 7.25), the label component was
regarded as lumps and non-road components.
Image data of roads was made through these processes for each
year’s photo. Changes of roads were visualized by comparing
these data. Although there still remains some errors and noises,
these images show change of roads clearly.
3.3 Results and discussion
The extracted road components of color aerial photos of 1984
and 1992 are shown in Figure 3 and Figure 4. By overlaying
these images, change of roads is visualized. As can be seen on
these figures, there still remains non-road component. By
comparing them with the original photos, it was found some
road components disappeared in the final result. Therefore this
method can be used for fast finding of change of roads but it
550
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
4.
CE
The
areas
situat
old n
aeria]
idea
enhar
effici
revisi
41V
We ;
const
