vertical roof height than residential buildings. Categorising the
building types using RMSE as discussed in section 4.1 to 4.3
therefore appears to be possible. Even though there is difference
in roof height between the two building types, could the
complexity of the roof structure be examined to a certain extent?
To investigate this idea further, the height difference between the
derived height from the 3D models constructed at various grid
resolutions, and a control height is determined. In this case, the
3D model is constructed using the maximum height derived from
the LIDAR DSM from 2m to 10m in grid resolution. For the
control height, a 3D model constructed using the 2m-grid
resolution LIDAR DSM was used. Four randomly selected
residential and industrial buildings are identified. Figure 15 and
Figure 16 shows the difference between the derived height and
the control height for the selected building using a grid resolution
of between 2m and 10m with 2m-grid interval for the residential
and industrial buildings respectively. There is a distinct difference
between the two plots corresponding to the residential and the
industrial buildings. From Figure 15, it is clear that the residential
buildings (a, b, c, and d) exhibit a high variation in height
(derived height — control height) as LIDAR DSM grid resolution
increases. This is due to the complex roof structure of residential
buildings, which have roofs of varying shape and height. For
industrial buildings (j, k, l, and m), the variation seems to be more
consistent beyond a grid resolution of 4m, as shown in Figure 16.
This effect might be due to the fact that the roof structures of
industrial buildings are less complex (i.e., exhibit less height
variation). Categorising building type by analysing the difference
between the derived height and the control height for each
building as illustrated in Figure 15 and Figure 16 seems possible.
Differences (m)
3 Building id.
2.5 ea
2 #6
1.5 OC
1 Ed
0.5
0
6
Grid Resolution .(m)
Figure 15: The difference between derived height and control
height for residential buildings.
Differences (m)
71 Building id.
6; €. — |
51 SN ok
| 3
A T NL el
31 :
21 TN
11
07
Grid Resolution (m)
Figure 16: The difference between derived maximum height and
control height for industrial buildings.
156
International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 3W14, La Jolla, CA, 9-11 Nov. 1999
4.5 Effect of Standard Deviation (Std. Dev.) on building
height
In the final part of the study, the variation of the Std. Dev.
measure for residential and the industrial buildings at various
LIDAR DSM grid resolutions is investigated. Figure 17 shows
the computed mean Std. Dev. for two samples of size 35 and 7 for
residential and industrial buildings respectively at various grid
resolutions. There is a distinct difference between the computed
mean Std. Dev. for residential and industrial buildings. The
computed mean Std. Dev. for residential building decreases
sharply from 1.6m to zero as grid resolution increases from 2m to
14m. This is partly because the surface area of residential
buildings is small, and so the value of the mean Std. Dev.
converges to zero as a result of averaging as the grid resolution
increases. However, for industrial buildings, the computed mean
Std. Dev. decreases steadily from 2m to 1.1m as grid resolution
increases from 2m to 16m and is almost constant beyond a 16m-
grid resolution. This is probably due to the large surface area of
industrial buildings, which gives a more consistent mean Std.
Dev. over this range of cell resolutions. Therefore, if the size of
the building is a significant factor in categorisation,
understanding the effect of mean Std. Dev. at various grid
resolutions appears to be a reasonable approach.
Mean Std. Dev. (m)
Grid Resolution
2m 4m 6m 8m 10m 12m 14m 16m 18m 20m
—e— Residential! 1.6 12 1 0.5 0.5 0.2 0 0 0 0
—e— Industrial 2 1.9 1.7 1.8 2 1.5 1.4 1.1 1.1 1.2
Figure 17: Mean Std. Dev. for residential and industrial building
at various grid resolutions.
Furthermore, since there is a distinct difference between the mean
Std. Dev. for residential and industrial buildings at various grid
resolutions, the ability of the Std. Dev. measure to identify
individual buildings is also investigated for randomly selected
buildings from each site.
Figure 18 shows the computed Std. Dev. for residential buildings
using grid resolutions ranging from 2m to 10m. The Std. Dev. for
each residential building (p, q, r and s) converges to zero at a grid
resolution of 10m. As mentioned earlier, this is due to the effects
of averaging. On the other hand, due to the complexity of the roof
shape of the residential buildings, significant variation in Std.
Dev. is seen at grid resolutions between 2m and 10m. Figure 19
depicts the computed Std. Dev. for the selected industrial building
(t, u, v and w) using grid resolutions of 2m to 10m. The computed
Std. Dev. values are almost constant in the industrial buildings
between grid resolutions of 2m to 10m. This is partly due to the
low-complexity of the roof structures as well as the greater
surface areas of industrial buildings. The behaviour of the Sid.
: Dev. for individual buildings appears to reveal properties of the
Internatic
nature of the roof ty
the two building type
Std. Dev. (m)
Figure 18: Std. De
and s) using 2r
Std. Dev. (m)
2.5
2 -—Q
1.5 B
1
0.5
0
Figure 19: Std. Dev
W) at 2m to
Differentiating betwe
using simple statistic:
possible. The findin
categorising resident
automated fashion. T
e By examining tl
LIDAR DSM gi
structures of re
inferred.
e The complexity
types can be e
individual buil
resolutions.
* The use of mear
models using th
RMSE comparec
* Categorising the
grid resolutions
be useful for cer
