an 
de- 
ys- 
to 
om 
de- 
of 
on. 
out 
cu- 
02) 
de- 
les 
e it 
ssi- 
the 
Ces 
The 
ess 
ork 
ited 
out 
tion 
sifi- 
and 
1age 
vel- 
| In- 
DA, 
International Archives of the Photogrammetry, Remote Se 
  
p=0.65 
   
‘agriculture —« 
   
p=0.05 
Figure 2: State Transition Diagram t — t+At, p = Probability for 
State Transition 
1 input image 1983 } 
  
J result 1983 | 
  
2 | input image 1988 |—- — result 1988 | 
3 result 1983 N classification } | result 1988 multitemporal - 
4 input image 2001 |.— ( 
= result 2001 | > 
  
5 result 1988 | pri *( classification 3 E 4 result 2001 multitemporal 
Figure 3: Data Flow Diagram 
the structure of the scene to be interpreted is modeled in a se- 
mantic network, allowing an effective hypothesis handling. The 
used state transition diagram is shown in figure 2. The arrows 
indicate assumed possible state transitions for regions between 
two images of different dates and its respective probabilities are 
placed near the arrows. Here inhabited area always stays inhab- 
ited area, the demolition of buildings and change to agriculture 
is very unusual. Changes from vegetation class to inhabited area 
are possible with probability of p — 3096. Additionally, changes 
from forest to agriculture and from agriculture to forest are pos- 
sible with probability of p — 594. 
In figure 3 the data flow is diagramed. First an aerial image of 
[983 is processed in the normal way, without any multitemporal 
knowledge. For the next step the interpretation result of 1983 is 
used as a priori temporal knowledge for the interpretation of the 
aerial image of 1988. The same is done with the input image 
of 2001. To assess the contribution of the multitemporal knowl- 
edge, the images of 1988 and 2001 are also interpreted without 
temporal knowledge. 
1245 
p=0.65 
nsing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
  
Figure 4: Segments for Classification 
3 CLASSIFICATION 
Two different approaches are followed to classify an input im- 
age. À region-based classifier works on segments of the image 
and classifies these segments into the considered 3 classes inhab- 
ited area, forest and agriculture. For an assumed application of a 
GIS verification system, the result is combined with a structural 
operator (see 3.2). The structural classifier searches directly for 
buildings that indicate the class inhabited area in the image. In 
the following section a description of both approaches is given. 
3.1 Region-based Classifier 
The region-based classification operator starts with a segmenta- 
tion of the entire image into segments of predetermined size. The 
size of a segment is chosen equivalent to that of an average house 
including a small garden. For each segment features are calcu- 
lated that are basis for the following linear regression classifier. 
The linear regression classifier uses a priori probabilities from 
the state transition diagram (see figure 2) with use of a previous 
classification. To get a more reliable interpretation the classifi- 
cation is repeated on a shifted segmentation of the image (see 
figure 4). For the black and green illustrated segments hypothesis 
are generated that overlap half of a segment size. The emerg- 
ing conflicts for one segment are solved by GEOAIDA, the most 
probable segment is taken as instance. The classification proce- 
dure is described in the next two sections. 
3.1.1 Feature Extraction Aerial images of different time are 
subjected to different illumination conditions and camera param- 
eters. The input images (see figure 1) are preprocessed by a con- 
trast stretch algorithm that unifies the color distribution in the his- 
togram for the following feature extraction operators.