queried in an attempt to derive roof ridge features from the
pixels. These features were then vectorised and edited. Small
error lines were removed and where possible ridge lines were
extended and attached to the original building outlines.
The algorithms were tested on buildings whose roof structure
was dominated by a main central ridge running parallel to
either of the key planimetric axes of the building (e.g. building
on left side of Figure 2). The assumption was that if the LIDAR
data could not extract the main ridge then it would not be able
to extract the finer elements of the roof structure. Some of the
industrial buildings were composed of multiple main ridges
and attempts were made to extract all these features.
3.3.1 Elevation. For each building the assumption made was
that the main ridge was the highest region of the building. The
higher the pixel value, the more likely that pixel holds ridge
information. The algorithm looked at the top 5-15% of roof
pixel values in an attempt to extract a continuous linear pixel
grouping of elevation values that represent the main ridge.
3.3.2 Slope. Around the main ridge, slope values change
abruptly, being close to zero at the ridge itself. The lowest 15-
30% of slope values were tested for their ridge extraction
performance.
3.3.3 Aspect. All buildings were assumed to be comprised of
two sloping roof segments joined by the main ridge. Aspect
values were split into two groups each representing one of the
two roof segments. The median aspect value was taken as the
threshold pixel value that separated the two groups. Values
below the threshold were classified as one and values above as
two. This created a border between the one and two pixels that
was extracted through vectorisation as the roof ridge.
4 RESULTS AND DISCUSSION
Statistics were produced from the assessment of the LIDAR
data set for roof ridge extraction. One statistic compared the
algorithm derived roof ridges with those collected in the field.
If a derived ridge was near parallel and close to the actual
building ridge then this was marked as a successful ridge
derivation. The decision as to whether a derived ridge matched
an actual ridge was made qualitatively from a visual
assessment. A percentage score was calculated for the number
of derived ridges that matched the actual ridges (Table 1).
A statistic was also produced that split the successfully derived
ridges into two groups. Those that exactly matched the actual
ridge in orientation and location were separated from those
classed as near matches. Table 2 shows the absolute number
and percentage of exactly matching derived ridges.
4.1 Algorithm Comparison
The aspect parameter derived from the gridded LIDAR data is
the most successful at roof ridge extraction (Table 1). It
performs better at both the residential and industrial sites.
International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 3W14, La Jolla, CA, 9-11 Nov. 1999
Derived/ Percentage Derived/ —
Actual Ridges : Actual Ridges
Parameter Residential Industrial Residential Industrial
Aspect 68/204 36/42 33.33 SS
Slope 50/204 22/42 24.51 82.38
Elevation 27/204 18/42 13.24 42.86
Table 1 Comparison of algorithm derived ridges for each
parameter to actual ridges.
Exact Match Derived Percentage of Derived
Ridges/Total Number of Ridges that Exactly Match
Derived Ridges Actual Ridges
Parameter Residential Industrial Residential Industrial
Aspect 35/68 28/36 51.47 77.78
Slope 47/51 22/22 92.16 100
Elevation 27/27 18/19 100.00 94.74
Table 2 Comparison of exactly matched derived ridges to total
number of derived ridges.
Slope is the next most effective parameter at ridge extraction
followed by elevation. The lowest 25% of slope values and top
10% of elevation values were found to be the most effective in
representing roof ridges. Figures 6a-e illustrate how each
parameter responds to an example building from the LIDAR
data set. Figure 6a is a 3D visualisation of a large; simple
roofed industrial building. It has two main roof sections, one
high and one low. Both sections are made up of two roof
segments that converge to a central ridge.
The performance of aspect is shown in Figure 6b. The two roof
segments can be clearly seen with a definite grey scale break
that defines the main ridge (Figure 6b.i). Both segments do
have some noisy pixels, but these are removed during the
execution of the aspect algorithm which groups all pixels
together for a particular roof segment. The extracted lines
based on the aspect values are shown in Figure 6b.ii. The main
ridge is successfully extracted on both levels of the building.
So too is the inverse ridge that marks the boundary between the
two roof sections. Several error lines can also be seen. These
are mainly due to the LIDAR sensor picking up the vertical
building sides and representing them as sloping roof segments
producing error ridges. This effect can be partly attributed to
the low scan angle of the LIDAR sensor distorting the building
shape.
For the slope and elevation algorithms, a certain range of
percentage values were extracted for each building to represent
the main ridge. This differs from the aspect algorithm which
focuses on homogenising the roof segments before the ridge
was extracted. The slope parameter manages to avoid the noisy
building edges because the slope values are high in those areas
and only the low slope values were queried in the algorithm.
This produces a cleaner representation of both main ridges.
The elevation parameter also avoids the noisy building edge
(Figure 6d). However, it is even less useful than the slope
Internation
algorithm as it only 1
ridges present on the |
hand side is lower
algorithm extracts hig
ridge, and therefore ¢
necessarily a proble
common for only one
areas multiple ridge b
b.i. Aspect value
c.i. Slope values
d.i. Elevation val
e.
Figure 6 Roo
Figure 6e is the a
processing stage, w
removed manually i
ridges extended wh
Structure. Error ridg
using the slope param
seem to cause the errc
Although the aspect
of roof ridges compai
exact matches is low
have near perfect rec
greater exploratory n:
is made between the
