International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
using the existing registrations in the ATKIS map data base (Wal-
ter, 2000). Experiments combining multi-spectral images (RGB,
colour infra red — CIR) with height information and reduction of
the information to generic surface types, have shown that it is
possible to perform automatic change detection with a satisfac-
tory accuracy (Petzold and Walter, 1999, Petzold, 2000) (note,
however, that the accuracy requirements for ATKIS are some-
what lower than for TOP10DK (Kort & Matrikelstyrelsen, 2001,
AVLBD, 1988)). The change detection leads to a “change map”
where the generic objects are divided in three classes: no change,
possible change and change.
In the Swiss ATOMI project, aerial colour photos, a high res-
olution Digital Elevation Model (DEM) and a Digital Surface
Model (DSM) are used aiming at the enhancement of the plani-
metric accuracy for the 2D VECTOR25 database (Eidenbenz et
al., 2000, Niederóst, 2003). The surface model is generated by
auto-correlation in aerial photographs in the scale of 1:10.000 and
is used as the primary data source. Image information (RGB/CIR)
is primarily used to discern man made objects from natural ob-
jects (buildings vs. vegetation).
Data from the digital multi spectral camera High Resolution Stereo
Camera—Airborne, HRSC-A (Neukum, 1999) is evaluated and
used within the Dutch project (Asperen, 1996, Hoffmann et al.,
2000). The HRSC-A data set includes high-resolution (15 cm)
spectral data (RGB and CIR) and an automatically generated
high resolution surface model from stereo matching,
A new change detection project within the framework of Eu-
roSDR is about to start up later this year. The emphasis is on de-
velopment of methods for localising changes in land cover from
very high resolution imagery, the integration of change maps in
the updating process and finally comparison of different methods
for change detection (EuroSDR, 2004).
2 DATA
The change detection procedure presented in section 3 below, is
evaluated using datasets mainly associated with the development
and updating of the Danish TOP10DK topographical map data
base.
2.1 RGB images
For the establishment and update of TOPIODK traditional RGB
aerial photographs have been used. All images are taken from an
altitude of approximately 3800 m leading to a scale of 1:25.000.
Each image covers an area of 6 km by 6 km and has a forward
lap of 60 percent and a side lap of 20 percent. As part of the
production work-flow the photos are scanned at a resolution of
21 um, leading to 350 MB of data, and a spatial pixel resolution
of 0.5 m at ground level. The photos were taken as part of a flight
campaign in April 2000.
2.2 Digital Surface Model (DSM)
As was described by Knudsen and Olsen (2003) it is very dif-
ficult to locate changes in the building layer using single aerial
images and hence only using spectral information in combina-
tion with size and form considerations. Therefore a high reso-
lution digital surface model (DSM) with a grid size of 1 meter
covering a test area in Lyngby, north of Copenhagen, has been
generated to facilitate the building detection. The dataset was
collected and made available for these studies by the Danish en-
gineering and mapping company COWL. Data were collected in
570
May-June 2001 using the TOPOSYSI system (Toposys, 2004,
Baltsavias, 1999) which only record first responses of the pulse.
The expected height accuracy is approximately 0.15 m.
2.3 Digital Map Database
The building theme from TOP10DK has been selected as target
for the update procedure. TOPIODK is a fully 3D map database,
including 51 object types (building, lake, highway ...) organised
in 8 classes (traffic, water ...). The precision of the database is
better than 1 meter both horizontally and vertically. For change
detection in the building layer, only new buildings larger than 25
m? and changes of building size larger than 10 m? are considered.
3 METHOD
The method presented is a revision of a method described by
Olsen et al. (2002) and Knudsen and Olsen (2003). It is based on
classification principles, using existing object registrations in the
map database as training areas in order to determine the charac-
teristics of the different classes used to search for and build the
object model. As it is very difficult to generate an unambiguous
object model for buildings using only spectral information, the
revised method also incorporates height information in the form
of high resolution DSM data e.g. from LIDAR or photogrammet-
ric auto-correlation to distinguish between objects in terrain from
objects above terrain.
3.1 The method step by step
The method which consists of three steps, preparation, classifi-
cation, and detection is outlined in figure 2.
Two major assumptions have to be fulfilled for the change detec-
tion procedure to be successful:
(1) The number of changes in a given class (e.g. building) must
be much smaller than the number of objects used to describe the
class. This is valid for most urban areas.
(2) New objects must share the same spectral characteristics as
the existing objects used to generate the object model. This is
often the case as only a small number of roofing materials is in
common use.
3.1.1 Preparation: The preparation consists of a data fusion
step to bring the data sets into a common reference frame and a
preprocessing step where various enhancement methods are ap-
plied to the data data to prepare them for the change detection
procedure.
Data fusion: as objects from the existing digital map database
is to be used as training areas for the determination of the class
characteristics. image data (raster) and the map database (vector)
must be co-registered. Generally co-registration can be done el
ther by registration of the image data to the map database or by
registration of the map database to the image data.
The most used method is registration of image data to the map
database. Howeyer the method has the disadvantage that most
image data types (aerial photos) have to be re-sampled as rectified
images or orthophotos. For the data sets to fit completely to each
other a high precision elevation model, including description of
man made objects (buildings, bridges etc.) must be available (i.e.
a Digital Surface Model. DSM).
Inte:
——
Fi
tic
