International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
. the object-based classification 
A good description of these data can be found in Janssen 
(1994). However, we can use the boundaries as certain 
knowledge but some of the crop types have changed. Using the 
available knowledge we update the crop types on this basic 
assumption that the boundaries are fixed during the time. We 
are also aware of the existing errors in the data and they 
influence the final results, as it will be shown in the following 
sections. 
3. GENERATION OF LIKELIHOOD MAPS 
The first step of our method is generating the likelihood maps. 
This was done in Idrisi 3.2 for Windows using BAYCLASS soft 
classifier. For this purpose we considered 7 spectral classes 
including: Beans, Cereals, Grass, Onions, Peas, Potatoes, Sugar 
beets. Training samples of these classes were defined using the 
GIS data and a false color composite of the 3 Landsat TM bands 
(bands 3, 4, and 5). Training for some of the classes was very 
difficult due to the similarity in spectral properties. For 
example, Cereals and Sugar beets are very similar in the 
generated color composite or the Grass fields and Peas fields 
are also hardly distinguishable. Then, we generated 7 likelihood 
maps for the spectral classes. Figure 3 shows two likelihood 
maps for Beans and Potatoes. 
In order to compare the results of the proposed method we 
performed a MLH classification on the TM data using the 
mentioned signatures. We exclude the area outside the GIS 
known boundaries from the classified image in order to have a 
better comparison of the results. Then all of the operations are 
only applied in the known regions and indeed the final results 
are also presented for this area. The MLH result can be seen in 
the Figure 2(c). 
4. IMPLEMENTATION 
In this method we use the agricultural fields stored in a GIS, for 
extracting more accurate results and finally using the obtained 
results we can update our GIS database. The proposed 
algorithm is on the basis of the generating the hypothesis maps 
and choosing the best hypothesis using the estimated costs. This 
leads to make use of the advantages of the top down approach 
in order to find the best results. Like the last example this 
procedure also works only for one class at a time and after the 
generation of the most accurate hypothesis or the hypothesis 
with the minimum cost the procedure is stopped. The procedure 
must be repeated for all the interested land cover classes. Figure 
4 shows a schematic diagram of the algorithm. 
In this experiment we use only the field boundaries from the 
GIS and ignore the other data stored in the GIS as crop rotation, 
crop calendar etc., which they can also be used in the method of 
MBIA. We assume that the agricultural fields have fixed 
boundaries and only the crop type of the polygons may change. 
This assumption in the modern agricultural areas such as 
Biddinghuizen region is usually correct and thus we can use it 
as a basic knowledge to extract the thematic information from 
the remotely sensed data. 
As it was mentioned in the last section, we generated the 
likelihood maps and stored them to be used in the next steps. 
Here the approach to integrate the RS and GIS data is simple. 
For each polygon we compute the average of the probabilities 
that are laid in it. Several statistical factors for the polygon like 
1283 
maximum or minimum value, average of values, sum of them 
and standard deviation of the grey values can be calculated. For 
our purpose we calculate the average probability for each 
polygon. 
Therefore, we calculate 7 average maps. Each of these maps is 
for an individual class. After this stage we take an average map 
for a certain class, for example Beans, and using a simple 
thresholding we can divide the average map into two classes, 
class Beans and Not-beans. The pixel, which has an average 
value greater than the threshold t, is assigned to the class Beans. 
The other pixels are labelled as the other class viz Not-beans. In 
fact, in this manner we generate a binary map in which the 
pixels that have the value 1 belong to class Beans and the other 
pixels with values 0 implies the inexistence to class Beans. 
Now we must test the selected threshold t. For this purpose we 
use a cost function (Abkar 1999) as below. It is defined as 
Cost = El H2 +E2H1 (1) 
We write the cost function for two classes because by working 
in the level of one class in each iteration really we have two 
classes: class x and others. Thresholding the average map 
generates the hypothesis map for class x (H1). Thus we can 
calculate the hypothesis map for the others such as below 
H2 zI-HI (2) 
in which I is a matrix that all of the its elements are |. 
  
[J Beans 
Cereals 
Grass 
  
L 
5 Onions 
Peas 
EM Potatoes