, CA, 9-11 Nov. 1999
Percentage Derived/
Actual Ridges
trial Residential Industrial
42 33.33 85.71
42 24.51 52.38
42 13.24 42.86
'orithm derived ridges for each
o actual ridges.
rived Percentage of Derived
ber of Ridges that Exactly Match
es Actual Ridges
ustrial Residential Industrial
8/36 51.47 77.78
2/22 92.16 100
8/19 100.00 94.74
ly matched derived ridges to total
derived ridges.
ve parameter at ridge extraction
/est 2596 of slope values and top
found to be the most effective in
ures 6a-e illustrate how each
mple building from the LIDAR
visualisation of a large; simple
ias two main roof sections, one
ions are made up of two roof
1tral ridge.
hown in Figure 6b. The two roof
with a definite grey scale break
Figure 6b.i). Both segments do
these are removed during the
rithm which groups all pixels
f segment. The extracted lines
shown in Figure 6b.ii. The main
on both levels of the building.
marks the boundary between the
or lines can also be seen. These
- sensor picking up the vertical
e them as sloping roof segments
ffect can be partly attributed to
\R sensor distorting the building
algorithms, a certain range of
ed for each building to represent
rom the aspect algorithm which
roof segments before the ridge
neter manages to avoid the noisy
pe values are high in those areas
; were queried in the algorithm.
entation of both main ridges.
avoids the noisy building edge
ven less useful than the slope
International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 3W14, La Jolla, CA, 9-11 Nov. 1999
algorithm as it only manages to identify one of the two main
ridges present on the roof. This is because the ridge to the right
hand side is lower than the left hand one. The elevation
algorithm extracts higher elevation values to represent the roof
ridge, and therefore any lower ridges are omitted. This is not
necessarily a problem in the residential areas where it is
common for only one main ridge to be present, but in industrial
areas multiple ridge buildings are much more widespread.
a. 3D Building Visualisation
b.ii. Aspect results
c.ii. Slope results
b.i. Aspect values
c.i. Slope values
d.ii. Elevation results
d.i. Elevation values
e. Edited Aspect results
Figure 6 Roof ridge extraction results for an industrial
building.
Figure 6e is the aspect parameter results after the final
processing stage, with the extracted ridges having been
removed manually if erroneous, and the correctly defined
ridges extended where possible to the original building
structure. Error ridges could be removed automatically by
using the slope parameter to eliminate the steep wall pixels that
seem to cause the error ridge problem.
Although the aspect parameter recognises the highest number
of roof ridges compared to slope and elevation, the number of
exact matches is lower than the other two parameters which
have near perfect records (Table 2). This is probably due to the
greater exploratory nature of the aspect algorithm. A trade off
is made between the tidier slope and elevation parameters that
are more likely to define a ridge in its exact position compared
to the aspect parameter that produces more error ridges, but
provides a more comprehensive absolute coverage of main
building ridges. The aspect parameter should still be preferred
regardless of the accuracy with which roof ridges are
represented because of its higher performance at ridge
recognition.
4.2 Site Comparison
The algorithm comparison provided an insight into which
derived parameters from the LIDAR data could be used to
extract roof detail. A comparison of the industrial and
residential area results in Table 1 and Table 2 highlight
properties of the LIDAR data that may affect the effectiveness
of the elevation, slope and aspect parameters at roof ridge
extraction.
The percentage values for derived ridges to actual ridges are
much higher in the industrial area compared to the residential
area (Table 1). The large buildings and simple roof structures
are largely responsible for this high rate of extraction success.
The LIDAR data set used in this paper has a grid resolution of
2m. For residential buildings this can cause problems, because
the pixel size is too large to adequately represent the
complicated variations in roof structure that residential
buildings generally display. Dormers, chimneys, television
aerials and other roof objects can all contribute to the
scrambling of roof ridge information held by the pixels.
Chimneys and other extrusions were also present on some of
the industrial building roofs but are small in comparison to the
roof size. They did not therefore affect the roof ridge extraction
to the same extent as in the residential area.
Another key difference between the two areas was the amount
of noise from objects surrounding the buildings. In the
residential area, neighbouring houses, trees, cars, hedges and
fences are often positioned close to a building. Information
from these features can mix with the building data as a result of
the interpolation of the LIDAR point data to a grid. This makes
the extraction of roof detail much harder. Mixed pixel effects
are not as prolific in the industrial area especially since
neighbouring buildings are positioned some distance from each
other, and most other objects are small in comparison to the
building itself.
Ground information can contribute to mixed pixel effects and is
a common problem for both areas. This results in building edge
information being smoothed. For small buildings ground
information can penetrate all the way into the central building
pixels and affect the extraction of roof ridges.
4.3 Other Influences on Results
Mention has already been made of the effect of the LIDAR
data’s grid resolution and mixed pixel effects on the results.
Other factors have also influenced the results in Table 1 and
Table 2. For this paper, the vertical accuracy of the LIDAR
data was found to have an RMSE value of about +/-0.3m. This
