International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
the homogeneity of the intensity and the average edge length
are derived. These attributes are combined in a classification
using fuzzy logic. 9096 correct classifications are reported.
Voegtle and Steinle (2003) also use a region merging
algorithm and a subsequent fuzzy logic classification. As
attributes of the segments they use the gradients on the
segment borders, the differences between first and last pulse
laser data and shape and height texture measures. With three
classes (bare Earth, building, and vegetation) 9376 correct
classifications are obtained.
Less results have been published on the actual change
detection using laser scanning data. Murakami et al. (1998,
1999) extracted changed buildings from multi-epoch laser
scanning data. Changed segments were delineated in an
image created by subtracting two images with digital surface
models. Steinle et al. (1999) compare laser scanning data
with an existing 3D CAD model of an urban environment. By
point wise comparison of heights changes can be seen.
Although no attempt is made to automatically detect the
changes, the potential of laser scanning data for this purpose
is clearly demonstrated.
Recently, Matikainen et al. (2003) presented a study on
change detection which compares classified segments of laser
data to buildings of a map. The comparison was performed
with a rule based system. A building was considered to be
recognised if e.g. 70% of the area of the building in the map
was covered by laser data that was classified as building
points. With a point density of 2-3 points/m? 91% of
buildings larger than 200 m“ and 42% of buildings smaller
than 200 m° were correctly recognised.
3. CLASSIFICATION
The extraction of the building segments from the laser
scanning data is performed in two classification steps. First,
the points are classified as bare Earth points or object points.
Next, the object points are classified as building points or
vegetation points. Both classification steps are performed on
segmented laser point clouds.
The separation of the bare Earth points from the other points
is performed with the algorithm described in (Sithole and
Vosselman 2003). The point cloud is divided into sets of
parallel thin slices in the XY-plane. The points of each slice
are considered as a profile. A minimum spanning tree is
computed for each profile. By removing the tree edges that
exceed a certain slope or length threshold, the minimum
spanning tree is split into line segments. All profiles are thus
segmented. This procedure is repeated for other sets of
profiles running in different orientations in the XY-plane.
Next, the resulting line segments of the different orientations
are merged. to surface segments. Two line segments of
different orientations are joined if they contain a common
laser point. The surfaces that are created have height
discontinuities all around their contours.
An advantage of this segmentation approach is that it is able
to deal with multiple overlapping surfaces. Thus layers of
vegetation as well as bare Earth points below this vegetation
can both be captured in segments.
The surface segments are classified based on the sign of the
height discontinuities at the ends of all line segments of a
segment. Only segments with a low proportion of line
segments that are above neighbouring line segments are
classified as bare Earth.
The remaining object segments are then further classified as
building or vegetation based on the values of one or more of
the following attributes:
e Surface roughness. Planes are fit to the points in small
neighbourhoods around each point of a segment. The
median of the standard deviations of all plane fits is used
as a measure for surface roughness.
e Segment size and height. A minimum segment size and
a minimum height above ground level can be specified to
select potential building segments.
e Colour (if available). Most providers of laser scanning
services nowadays offer the simultaneous recording of
imagery. When registered, a colour value can be assigned
to the laser points by projecting the points into the
imagery and interpolating the colour value. In particular
the hue value of colour imagery can be used to
distinguish vegetation from most roof materials, but also
the intensity value proved to be useful. Median values
can be computed for the laser points within each
segment.
e First-last pulse difference (if available). The difference
between the first and last pulse recording is known to
give a good indication for the presence of vegetation
(Oude Elberink and Maas 2000). Although large
differences can also be observed at the edges of
buildings, the median value of the height differences of
all points within a building segment should clearly be
lower than the medium value of the height differences
within a vegetation segment.
The different attributes are combined in a K-nearest
neighbour classification to obtain the classification for each
segment. After the classification, the building segments that
are adjacent in the XY-plane can be merged to form larger
segments. These segments should then correspond to
complete buildings.
4. CLASSIFICATION RESULTS
The above classification method was applied to laser scanner
data of (a part of) the city centre of Nijmegen. The data was
recorded with an Optech ALTMI225 scanner with an
average point spacing of 1.2 m. Colour imagery was recorded
simultaneously. The result of the classification is shown in
Figure | and quantified in Tables I and 2. The segment-based
filter showed no problems in removing larger buildings, a
well-known problem for morphological filters (Sithole and
Vosselman 2004).
The separation of buildings and vegetation was performed
using the roughness and colour information of the segments.
Compared to manually classified data used as ground truth,
85 % of the building points and 78 % of the vegetation points
were classified correct. The overall classification accuracy
over the three classes bare Earth, buildings and vegetation
was 90%. The ground truth of those points that were
classified incorrectly is shown in Figure 2. Several kind of
errors can be observed in this figure:
Intern
Grou
truth
Bare
Builc
Vege
| Total
Grou
truth
Bare
Builc
Vege
