IX 
International Archives of the Photogrammetry, Remote Sensin 
g and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
  
0.184 [- \ / 
| | | | | 
: 8 esnold : 
O15 2. 632. 044. IY ire 
  
  
  
Figure 5. The cost function of the hypotheses maps for class 
Onions as a function of threshold t. t values have step of 0.1. 
  
  
  
  
  
  
  
  
  
  
  
  
Class in o Minimum Cost 
1 Beans 0.50 0.071903 
2 Cereals 0.49 0.116372 
3 Grass 0.50 0.088656 
4 Onion 0.50 0.182127 
5 Peas 0.5 0.0302247 
6| Potatoes 0.51 0.085219 
7| | Sugar-beets 0.5 0.150371 
  
  
Table 1. The best threshold value and corresponding minimum 
cost for each class 
As it is clear most of these polygons are not consist of a single 
crop type (the basic assumption that we assumed). This shows 
that the proposed method is sensitive to some common errors 
that many of the usual methods ignore them. Therefore, we can 
be sure that if we assume the fixed boundaries, this is a true 
assumption and if this assumption violates for a specific 
polygon then this polygon will not be labelled. Thus, unlabeled 
polygons often are which within them some new internal 
boundaries have been produced and the polygon has been split 
into the new sub fields. Of course, some of these unclassified 
polygons have a single crop type and the other errors have 
caused the polygon not to be labelled. One of the major sources 
of these errors is the existing error in GIS data. These caused 
that the polygons are not matched at their place exactly and 
indeed some pixels which have high probability values are not 
fall into those polygons. Some examples of such polygons have 
been shown in the Figure 7. 
5. OBJECT-BASED CLASSIFICATION 
A common approach for classification of the remotely sensed 
images is Object-based classification (OBC) (boundary-based or 
parcel-based classification) using the boundaries in a GIS. In 
this method a traditional classification like MLH, is done and 
each pixel is assigned to the most probable class. After this, the 
existing boundary map is overlaid on the classified map. For 
each polygon the class, which is the more frequent or has the 
majority in that specific polygon is assigned to it. In such a 
    
method usually all of the polygons or fields are allocated to a 
class because often there is a majority for one class in the 
polygons. We expect that this method should give the improved 
result relative to the traditional MLH approach. But as we will 
see in the next section, it dose not give the required information. 
This method is based on a simple logic that if one land cover 
class is in a field (in the real world) indeed it must have the 
majority for that (in the classified map). But this fails in the 
small fields relative to the sensor spatial resolution because of 
the radiometric overlaps between the classes and also errors, 
which are involved in the MLH classification. Consequently, 
we can see that this method cannot give results as good as the 
MBIA approach. The result of this method has been shown in 
Figure 6(a). 
Grass 
Onions 
Peas 
Potatoes 
B] Sugar-beets 
      
E d 
Figure 6 Result of the MBIA and OBC approaches (see the 
various crop types in the superimposed polygons). 
a) final result of the OBC b) final result of the MBIA 
¢) undefined polygons by the MBIA superimposed on the color 
composite of the 3 TM bands 
6. EVALUATION OF THE RESULTS 
So far, we classified the Biddinghuizen RS data in three ways: 
Maximum Likelihood classification, Object-based classification 
1285