International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
with NDVI > 0.1. The result of the validation is a refined build-
ing mask, holding only the buildings which are most likely still
buildings.
3.1.2. Classification: The first step in the classification part is
to perform a clustering process. As stated by e.g Kressler and
Steinnocher (1996) some classes, (e.g. buildings) have to be sub-
divided into more unique subclasses as they are spectrally highly
diverse. This task is handled by splitting up the group of pix-
els registered as buildings by the building mask into smaller and
more unique sub-classes using a simple migrating means cluster-
ing process. The algorithm is based on the ISODATA algorithm
(Ball and Hall, 1965), and the number of sub-classes is automat-
ically determined by the algorithm in order to make a best fit to
the input dataset.
The clustering process is followed by an actual classification.
The sub-classes, which are spectrally more uniform than the base
building class, are used (either alone or in combination with other
class descriptions (e.g. water, roads, forest, grassland etc.) to
perform a Mahalanobis classification of the entire image. This
causes all pixels in the image to be assigned to the class hav-
ing the smallest Mahalanobis distance from the pixel value to the
class (Richards and Jia, 1999). Threshold values all being depen-
dent on the class characteristics, are used to assign pixels with a
distance too far from the closest class to a garbage class.
The two successive steps are run a number of times as part of
an iteration process. This is done mainly due to the fact that the
result of the ISODATA algorithm is strongly dependent on the po-
sition of the initial cluster centres. After this iteration process it is
possible to accept pixels identified as buildings a specific number
of times. In the case study presented below all pixels which are
classified as a building one or more times are considered to be
a “building”, leading to an image holding pixels with values of
either zero or one: zero indicates no building and one indicates a
potential building pixel.
Using the nDSM, the image holding potential buildings are fil-
tered in order to extract only objects (pixels) which stands above
terrain.
3.1.3 Change detection: First a change map is computed by
a pixel by pixel comparison of the existing map database (in a
raster version) to the classification result. Since the change map
includes all potential changes in the building layer it includes
noise in the form of single pixels, and some false alarms due
to misclassification. The single pixels are removed using mor-
phological opening. The remaining change pixels are segmented,
and pixel clusters smaller than the detection requirement (c.g. 25
m? = 25 pixels in the TOP10DK case) and/or not fulfilling the
size and shape specifications for buildings are removed from the
dataset, leading to a reduction of the false alarms and the final
change map.
4 CASE STUDY
The procedure is tested on the data described in section 2. The
latest update of the TOP10DK database was carried out five years
before the photos were taken.
4.1 Test area
The test area used for evaluation is situated in Kgs. Lyngby, a
suburb 15 km north of Copenhagen. The area contains many dif-
ferent types of buildings and houses since it includes a small in-
dustrial area; a cemetery; a church; a small train station; large
572
strip buildings; and a gasoline station. The looks and shapes of
the buildings as well as the heights differ a lot. Vegetated areas
take up a large part of the area, and since the area also includes a
highway, two bridges (one for pedestrians and one for cars) and a
rail road, this causes a very special terrain structure. The area is
also characterized by the fact that many changes have taken place
since the establishment of the TOPIODK database.
Area: Approximately 700m x 500m. Lower left corner (E, N) «
(718450,6187050). Upper right corner (E, N) = (719150. 6187550).
(E, N) coordinates are given in UTM zone 32. All images (RGB,
TOP10DK and DSM) are 500 rows by 700 columns and approx-
imately 70 buildings are included in the area. 72 registrations are
included in the existing map data base. 12 new houses have been
build since the last revision and 14 have been demolished.
4.2 Test data
All datasets are subsamples from larger datasets and they are
brought into the same geographical reference by orthorectifica-
tion of the aerial photographs using an existing digital terrain
model (DTM) with a grid spacing of 20 m.
4.3 Results
The results are visualised. in figure 4. The first image shows the
RGB image with the existing map database superimposed in yel-
low colour. It can be seen that a lot of development has taken
place since the last revision of the map database. This is most
pronounced in the right part of the image where 7 buildings have
been demolished and 10 new buildings with blue roofing material
have been built.
The second image from the top shows the result after the classi-
fication step. White pixels indicate potential buildings, and as it
can be seen large areas are misclassified as vegetation and roads
are classified as potential buildings.
The third image shows the result after the height filtering. It can
be seen that all roads are now removed. Some vegetated areas
still remain as potential buildings, though.
The last image shows the RGB image with the changes found by
the automatic change detection algorithm superimposed in yel-
low. The results are summarised in table 1.
Factual | Detected
Demolished Buildings 14 12
-NewBuüldnss | | (1 7$
Changes 26 14
False alarms 45
Table 1: Statistical results
Approximately 50 percent of the factual changes in the test area
have been detected by the algorithm.
5 DISCUSSION
Most success is experienced in the group of demolished buildings
where only two demolished buildings have not been detected.
This is caused by the method used for change detection where
the detected buildings are compared directly to the existing reg
istrations on a pixel wise basis. The two demolished buildings
which are not detected are positioned the upper right area of the
. Interna
A
