stanbul 2004 
AUTOMATIC CHANGE DETECTION FOR VALIDATION OF DIGITAL MAP DATABASES 
Brian Pilemann Olsen 
National Survey and Cadastre, Denmark 
8 Rentemestervej, DK-2400 Copenhagen NV, Denmark 
bpo@kms.dk 
Working Group IC WG II/TV 
KEY WORDS: Photogrammetry, Remote Sensing, Change Detection, Classification, Automation, High Resolution, Infra-red, 
DEM/DTM 
ABSTRACT 
In almost all areas of our society there is an increasing need for up to date digital map databases. Traditionally, different manual, 
labour intensive and hence costly methods have been used for map updating, with the change detection for the updating being by far 
the most complex and expensive part. In this paper an automatic change detection method is presented and evaluated. The method only 
considers changes in the buildings theme, but it can be extended to other object classes. The aim is the development of an efficient 
change detection procedure for database maintenance in a production environment. The method is based on classification principles 
and combines an unsupervised and a supervised classification in order to determine the spectral response of the building class and thus 
locate potential buildings. The result is filtered by a height filter to refine the result. The method is evaluated on building registrations 
from the Danish TOP10DK map database. The test case presented in the paper is from a residential suburban area. The method detects 
almost all changes due to demolished buildings whereas only a smaller part of the new buildings are detected. This is primarily due to 
the use of a very special roofing material. The method leads to a number of false alarms, which to a large degree can be eliminated by 
refinements of the algorithm or by introduction of additional information e.g. infra-red images or texture measures. 
1 INTRODUCTION 
The digital topographic map database TOP10DK is the primary 
topographic product of the National Survey and Cadastre—Den- 
mark (Kort & Matrikelstyrelsen, KMS).The development and up- 
date of TOP10DK is based on aerial photogrammetry: one fifth of 
Denmark is photographed each spring before foliation, resulting 
in approx. 1200 photos (about 400 GB of image data). 
Map updating can be carried out by a complete remapping of 
the area for each revision cycle, but much work can be saved by 
detecting changes from the previous version of the map database 
and concentrating on these areas of change. Change detection 
for topographical mapping is on the other hand not a simple task: 
Although the intention is always to carry out the photo flights 
at approximately the same time of year, the natural, inter annual 
variations of the vegetation coverage is of a magnitude that hides 
the (primarily human generated) changes sought for. Furthermore 
it is almost impossible to take the photos at the same geographical 
position and with the same attitude as within the previous photo 
campaign. This means that the change detection must be carried 
out by comparing a new image directly with the existing map 
database, rather than by a simpler image-to-image comparison. 
In this paper an automatic procedure for change detection con- 
centrating on buildings, which are important mapping objects, is 
presented. The next step in the update process: the actual 3D ob- 
ject registration, is not considered here. This subject has recently 
been treated extensively by Niederóst (2003) and Süweg (2003). 
As can be seen from figure 1 buildings are often highly diverse 
both when it comes to size, form and spectral signature. They are 
therefore hard to describe by spectral information only. Adding 
height information to the process, e.g in the form of digital sur- 
face models, may improve the distinguishing of buildings from 
other objects having a similar spectral response. 
When introducing automatic change detection procedures, the 
atm is to detect at least the same percentage of factual changes as 
569 
  
Figure 1: Buildings are typically highly diverse and spectrally 
ill-defined, when considered as a single group. The last image 
(Lower Right) is a building as it is represented in a DSM. 
a manual operator is capable of. It is, on the other hand, accept- 
able if the change detection procedure introduces false alarms, as 
long as they are few, since they can easily be rejected during the 
actual 3D object registration. 
1.1 Related work 
Other European countries e.g. Germany, Switzerland, and the 
Netherlands have also established and completed digital map data- 
bases with national coverage in the past few years. The National 
Mapping agencies in these countries therefore face the same prob- 
lem as the KMS and projects with similar aims considering auto- 
matic or semi-automatic map updating have been established. 
In Germany the project for updating the ATKIS database focuses 
on registration of more generic surface types (settlement, grass- 
land. street, water etc.) (Petzold and Walter, 1999, Walter and 
Fritsch, 2000). The method for change detection uses super- 
vised classification, with training areas automatically generated