dfgi.fi
n
data to be
etected by
trees and
re used in
buildings.
fected and
arger than
tween the
c building
Id include
nethod.
ation and
abase and
s could be
inner data
they have
modelling
1an, 1999;
Fujii and
he height
ther steps
t article 1s
al change
anner and
e building
lding map
similar to
rial ortho
and some
he method
er scanner
is been in
son of the
presented
>sults was
out 15-20
t al., 2004:
ser scanner
" about 0.4
issification
x detection
he training
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
area and to test areas covering about 2.2 km“. The accuracy of
building detection was estimated on the basis of the test arcas.
which can be roughly divided into three types of area with
different characteristics: an industrial area, an apartment house
area and a small-house area. Topography in the study area is
varying and characterized by small hills.
The laser scanner data were acquired with the TopoSys
FALCON system on 14 May 2003, when some trees were still
without leaves and others had small leaves. The flying altitude
was 400 m a.g.l., which resulted in a point density of about 10
points per m?. Due to an overlap between adjacent strips, the
average point density in the dataset is about 17 points per m^. A
digital surface model (DSM) in raster format was created from
the laser scanner data using the TerraScan software (Terrasolid,
2004). To each pixel, the highest value within the pixel in the
first pulse data was assigned, and interpolation was used to
determine values for pixels without laser points. The original
first pulse laser points were also classified in TerraScan to
detect points located over 2.5 m above the ground surface.
Ground points were first detected by a routine that iteratively
builds a triangulated surface model (Soininen, 2003). Using
another classification routine, other points were then compared
with a temporary surface model based on the ground points.
Classification of the points was used as a substitute for a digital
terrain model (DTM) to distinguish buildings and trees from the
ground surface in a later stage of the study (see Section 3.1).
An intensity image was also created from the laser scanner data.
The intensity value corresponding to the highest last pulse
height within the pixel was first assigned to each pixel (intensity
information was not available for first pulse data), and
interpolation was then applied. However, the intensity image
did not appear very useful and was not used for building
detection in the study. In overlap areas between different strips,
the image had a grainy appearance, probably due to differences
in intensity values between/across the strips.
Aerial colour imagery in a scale of 1:5300 were acquired and
scanned by FM-Kartta Oy. The images were taken on 26 June
2003. An ortho image was created with Z/I Imaging
ImageStation Base Rectifier (Z/I Imaging, 2004) using the
laser-derived DSM. Comparison of the rectified image with
reference data shows that buildings are accurately located.
However, it must be noted that areas behind buildings or trees
in the original imagery are not correctly presented. They are still
covered with the building roof or tree canopy, which reduces
the usefulness of the imagery for building detection.
Buildings of the Topographic Database of the National Land
Survey of Finland from 2000 were used as an old map to be
updated. A building map from 2003 obtained from the city of
Espoo was used as up-to-date reference data in rule
development and accuracy estimation. The positional accuracy
of objects in the Topographic Database is about 5 m (National
Land Survey of Finland, 2002). Visual comparison with other
data sources shows that most buildings in the study area are
accurately located. The building map from the city of Espoo
presents the buildings in more detail. Compared with some
ground measurements in the study area, the positional accuracy
of buildings in the map is 0.5 m or better. The map data were
converted from vector format to raster maps. From the reference
map, polygons smaller than 20 m’ were eliminated before
conversion to exclude very small buildings and other
constructions from accuracy estimation. On the other hand,
some smaller parts of larger buildings also became eliminated.
435
[t must also be noted that despite its accuracy, the building map
is a generalized representation of the buildings. Compared with
the laser scanner data and aerial imagery, many differences can
be observed. This must be accounted for when accuracy
estimates calculated on the basis of the map are investigated.
In addition to the building maps, a forest map obtained from
FM-Kartta Oy was used in the study. It was used in the training
area in developing classification rules for building detection.
The DSM, intensity image. aerial image and map data were all
processed into raster format with 30 cm x 30 em pixels.
3. METHODS
3.1 Building detection
The building detection method was based on the following
steps:
I. Segmentation of the DSM into homogeneous regions
2. Classification of the segments into two classes:
'ground' and 'building or tree', based on the classified
laser points (see Section 2)
Classification of 'building or tree' segments into
buildings and trees using height texture, aerial image
and shape of the segments
4. Improvement of the classification result using size of
the segments and neighbourhood information in
addition to the three attributes above
Classification-based ~~ segmentation to
neighbouring building segments
6. Classification of the new segments based on the
UG
CA
merge
previous classification result
Segmentation and classification, except step 2, were performed
using the eCognition software (Definiens Imaging, 2004). The
segmentation method of eCognition (Baatz and Schápe, 2000;
Definiens Imaging, 2003) is based on bottom-up region
merging and a local optimization process minimizing the
growth of a given heterogeneity criterion. A heterogeneity
criterion based completely on colour information, which in this
case corresponded to height in the DSM, was used.
The segments were first classified into two classes: 'ground' and
'building or tree', using the laser points classified in TerraScan.
This was conducted in Matlab (The MathWorks, 2004) by
calculating the number of points over and under 2.5 m above
the ground surface within each segment. Within each pixel, only
the highest point, which was also used in forming the DSM,
was considered. The segment was classified as ‘building or tree’
if most of the points had a height value over 2.5 m, otherwise as
'ground'. The classification result was imported into eCognition
as an additional image layer and used to classify segments into
‘ground’ and building or tree’.
Attributes for distinguishing buildings and trees from each other
were selected after investigating the histograms of known
building and tree segments in the training area (segment was
used as a training segment for building or tree if over 80% of it
belonged to building or forest in the map data). Attributes under
study included mean values and standard deviations of height,
intensity and aerial image channels, size, various shape
attributes and various texture attributes. The attributes were
exported from eCognition for analysis. Three attributes were
selected for classification: 1) Grey Level Co-occurrence Matrix
