anbul 2004 
osys, 2004, 
f the pulse. 
n. 
ed as target 
ip database, 
) organised 
database is 
For change 
rger than 25 
considered. 
escribed by 
is based on 
itions in the 
the charac- 
nd build the 
nambiguous 
mation, the 
in the form 
togrammet- 
terrain from 
on, classifi- 
ange detec- 
ilding) must 
describe the 
cteristics as 
del. This is 
aterials is in 
data fusion 
frame and a 
hods are ap- 
ge detection 
ap database 
of the class 
jase (vector) 
| be done ei- 
tabase or by 
1 to the map 
ge that most 
d as rectified 
etely to each 
escription of 
vailable (Le. 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
  
DATA SOURCES 
pata. || RE || MR) or 
BASE IMAGES IMAGES DSM 
  
  
  
  
  
  
  
  
  
  
: 
| 
| 
| | 
| PRE- | 
| 5 | PROCESSING 
i © | 
| © | 
IS | 
n s | 
LS | 
|. | 
i DATA FUSION i 
| 
| 
| neri ie ee => 
CLUSTERING 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
I 
| 
| 
I 
l | 
| | 
l | 
| | 
| | 
| | 
| | 
| | 
| o i | 
} 2 | 
5 | 
; 2 CLASSIFICATION | 
8 | 
| 3 | 
i l 
I l 
i 
| | 
| | 
| | 
I | 
| | 
| | 
| | 
waa em ] 
TE J 
| | 
| | 
| | 
CREATE 
| 5 CHANGE MAP | 
| 8 Y I 
| o I 
| == POST 
| PROCESSING | 
! 
| i 
| | 
fe 
Figure 2: Change detection work flow—<cf. section 3 for descrip- 
tion 
571 
Another way is to project the map database directly to the other 
data sets, e.g. onto the aerial images using the basic photogram- 
metric equations (Kraus, 1993). For this to work a database with 
(X, Y, Z) coordinates and the orientation parameters for the aerial 
images have to be available. This method, leads to the most pre- 
cise co-registration, and eliminates any resampling of image data. 
Preprocessing: Various algorithms are applied to the data set to 
prepare them for the change detection process. The three most 
important processes are: (1) calculation of NDVI (Normalised 
Difference Vegetation Index) images if colour infra-red (CIR) 
photos are available; (2) generation of a normalized Digital Sur- 
face Model (nDSM); and (3) evaluation of training areas. 
NDVI is calculated as NDVI — ir-1e0; NDVI is well suited for 
distinguishing vegetated areas from man made objects. 
A nDSM only includes objects which stands above terrain and 
it can be calculated as using a Digital Terrain Model (DTM): 
nDSM = DSM — DTM. If a DTM is not available it must be 
estimated from the DSM. A very simple method for DTM esti- 
mation using grey tone morphology is described by Weidner and 
Fórstner (1995), and used in these tests. First a minimum filtering 
of the DSM is performed using a flat structuring element B (with 
a given size and form). In this way the minimum height in the 
area determined by the structuring element is assigned to the ori- 
gin of the structuring element (pixel). This minimum filtering is 
followed by a maximum filtering, using the same flat structuring 
element. Performing the two steps in the described order equals 
an morphological opening: z = z o B and leads to an estimation 
or approximation of the topographic surface, the DTM. In order 
to eliminate all elements above terrain (buildings), the size of the 
structuring element must be chosen in such a way that it is not 
completely contained in a building. The size depends on the area 
to be processed. If a priori information concerning existing build- 
ing sizes in the area is available the size can be fixed using this 
information. In the test presented in this paper the size of B is 
fixed to 25 m. The process is illustrated in figure 3. 
  
   
  
  
  
  
  
  
  
  
55 55 25 
50 So 20 
45 45 ; 
1 
40 A 40 
10 
35 35 
30 30 — $ 
25 25 ol I 
50 100 150 50 100 150 50 100 150 
Figure 3: nDSM creation using artificial DTM. UL: DSM. UM: 
estimated DTM, Z = zo B. UR: nDSM = DSM - DTM. LL: DSM 
profile. LM: DTM profile. LR: nDSM profile. All profiles follow 
the red line in the DSM, DTM and nDSM respectively. 
Validation of the training areas is done using the estimated nDSM 
and/or the NDVI image. An objects above terrain mask can be 
generated using a height threshold of e.g. z < 2.5 meters in the 
nDSM. With this mask, areas registered as buildings in the ex- 
isting map database, which no longer stand above terrain are fil- 
tered out. Objects covered by vegetation can be eliminated using 
the NDVI mask (if available), as they can be detected as areas