International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
On the other hand, trees growing beside and sometimes partly
over buildings often became connected into same segments with
buildings, which enlarged the buildings in the classification
result and lowered object accuracy.
Table 2. Building-based accuracy estimates showing the
percentage of buildings correctly detected in
building detection (all test areas included).
Buildings of the reference map
Building | Percentage | Minimum Total Buildings
size threshold | member- | number of | correctly
» ship buildings | detected
All 70% — 813 87.6%
>200m’ | 70% = 226 97.8%
<200m | 70% - 587 83.6%
50% — 587 85.7%
Buildings of the classification result
All 70% = 791 58.2%
0.75 578 70.8%
50% — 791 81.2%
0.75 578 95.7%
>200m“ | 70% - 333 70.9%
0.75 306 74.2%
50% = 333 96.4%
0.75 306 99.0%
«200m' | 70% - 458 48.9%
0.75 272 66.9%
50% = 458 70.1%
0.75 272 91.9%
*) Percentage threshold shows the required overlap for
buildings of the map and buildings of the classification result.
The building-based accuracy estimates show that 87.6% of
buildings in the map were detected when an overlap of 70%
with classified buildings was required. For buildings over 200
m?, the detection percentage was 97.8%, and for buildings
under 200 m’, it was 83.6%. This can be considered as a
satisfactory result. Visual evaluation of buildings not detected
shows that many of them are not visible or not clearly visible in
the laser scanner and aerial image data, e.g. due to trees. Some
buildings presented in the map are also lower than 2.5 m, which
was used as a threshold value in classification. For buildings of
the classification result (lower part of Table 2), the required
overlap with buildings of the map had a large influence on the
accuracy estimates. This is related to the larger building size in
the classification result than in the map. With an overlap
requirement of 70%, 58.2% of all detected buildings were
correct buildings. When the overlap requirement was decreased
to 50%, the percentage of correct buildings increased to 81.2%.
As expected, large buildings were correct buildings more
probably than small ones. The results also clearly indicate that
the membership value to building from classification provides
useful information on the reliability of the detected building. Of
all certainly detected buildings (membership over 0.75), 95.7%
with an overlap requirement of 50% were real buildings.
It can be concluded that good building detection accuracy was
obtained, which is important for automated map updating. The
results also indicate that the positional accuracy of detected
buildings compared with the reference map was not perfectly
good, which is partly related to representation of buildings in
the map and characteristics of the data. Some errors in building
detection, e.g. connection of buildings with trees, also occurred.
438
The results are in accordance with results from another study
area and dataset (laser scanner data with lower pulse density, no
aerial imagery). In that study (Matikainen et al., 2003), an
interpretation accuracy of 90.0% and an object accuracy of
85.4% were achieved. About 80% of all buildings and about
90% of buildings larger than 200 m? were detected.
4.2 Change detection
Change detection results for the entire test areas are shown in
the lower part of Figure | and for a selected subarea also in a
larger scale in Figure 2. On the basis of visual evaluation, the
following conclusions can be drawn:
— Objects classified as certainly detected new buildings
were normally new buildings or building-like
constructions.
— Objects classified as uncertainly detected new
buildings were typically misclassifications.
— Of 19 major new buildings in the study area, 17 were
detected as new buildings with certain detection, 2
were partly classified as tree and partly as new
building with uncertain detection.
— Many buildings classified as enlarged or partly
detected were presented differently in the old and new
maps. The classification in these cases can be
considered correct.
— Many buildings were also classified as enlarged
because they were connected into same segments with
trees and/or because they appeared larger in the data
sources than in the map. Some buildings were partly
classified as tree and thus labelled as partly detected in
change detection.
— More advanced rules for detecting enlarged buildings
should be developed. In some cases an enlargement of
a building was correctly labelled as building in
building detection, but in change detection the
building was classified as an old building due to the
small size of the enlargement compared with the size
of the building.
— Buildings classified as not detected were typically
small buildings difficult to detect or two-level car
parks. Many of the car parks are located on a slope
with one side of the upper level on or near the ground
surface and thus easily became classified as ground
(e.g. the building in the upper left corner of Figure 2).
— Some of the not detected old buildings did not exist in
the reference map, i.e. they were correctly classified in
change detection.
Development of the change detection method is still in an initial
stage, but as described above, promising results were obtained,
especially in detection of new buildings, which is the most
important task for map updating.
5. CONCLUSIONS
Automatic building detection and change detection from laser
scanner and aerial image data was studied. Good building
detection accuracy was achieved, which was also the main goal
of the study. The interpretation accuracy of buildings (pixel-
based estimation) was 9494. According to building-based
accuracy estimates, 88% of all buildings and 98% of buildings
larger than 200 m^ were correctly detected. Further research
should include improvement of the segmentation stage to better
distinguish buildings from trees (e.g. use of aerial imagery with
Interna
visible
laser s
metho«
the ma
the res
already
detect:
updatlı
Ahoka:
assessr
Submit
Remot
Turkey
Armen
2003.
mappit
Remot
Baatz,
optimi
segmel
Geogr.
AGIT-.
12-23.
Baltsa
analysi
and st
Photog
Defink
Analys
Germa
Defink
Fujii, |
airborr
Transa
2234-2
Haala,
in. urb.
Remot
Hoffm
Towar
data-ac
Archiv
The Ne
Jung, |
aerial :
Sensin,
Knuds
for u
Engine
Maas,
extract
Journe
163.
