rs of 
the 
| the 
n be 
seed 
rder 
onal 
rion 
rder 
rid. 
> 
s Ç 
(1) 
ipes 
y be 
uch 
> of 
ber 
der 
r of 
are 
nth 
lers 
the 
| as 
the 
ent 
ork 
the 
Olaf Hellwich 
  
In the first case a high-resolution (1m pixel size) panchromatic image (Fig. 6 a)) is combined with a multispectral image 
(Fig. 6b), 4m pixel size) and a digital surface elevation model (DEM) with 1 m grid size. The imaged area is located 
in Upper Bavaria and consists of agricultural fields, patches of forest, and rural residential areas. The road network was 
extracted automatically from the high-resolution image starting in open areas detected by landuse classification. The 
multispectral data was analyzed using unsupervised clustering and classified into 15 landuse classes with an algorithm 
including a Markov random field model supporting the detection of continuous areas belonging to a single class. Ambi- 
guities between non-vegetated areas and buildings were solved with the help of a high-objects class extracted from the 
DEM (Fig. 6c)). Edges from the high-resolution image were combined with the road network in order to extract agri- 
cultural field units. Inside of an agricultural field a majority filter was applied to the classification of the multispectral 
data selecting the mode of the classes occuring in a field as the landuse of the complete field. The resulting classification 
(Fig. 6d)) was compared with a limited set of ground truth data. Of the eight agricultural fields with known crop one 
summer wheat field was misclassified as winter wheat. The classification procedure is described in detail in (Hellwich 
and Wiedemann, 2000). In addition to the usefulness of multisensor fusion, this example demonstrates that also the com- 
bined extraction of different object types, here agricultural fields and roads, improves the accuracy and reliability of the 
resulting interpretation of the scene. 
meadow 
C) : d) 
  
Figure 6: a) High-res. panchromatic image, b) IR color image, ¢) high objects (dark) extracted from DEM, d) landuse 
classification (with imprinted ground truth information). 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 393