134 
JUL GflMMR IN DB JUL 1600 GOMMA IN DB 
- 
T 
1 
R 
2 
U 
3 
E 
4 
L 
5 
6 
A 
7 
B 
8 
E 
9 
L 
10 
Table 2. C 
segmentati 
of labels) 
Figure 6. Feature space plots of May vs July and the 2 July features (low and high altitude). 
are rather small. Even the winterwheat seems to be 
separable in this plot, but since this can be done in 
May, no further attention is paid to it. To reduce the 
dataset, we now introduce a linear combination deter 
mined as the first eigen vector of the two July 
features, to optimise the separability of the crops. 
Figure 7 shows the plot of May versus the combina 
tion of the July features. Figure 8 shows a histogram 
of the July data projected on the new axis. Winter- 
wheat fields are excluded in this histogram. The peaks 
are from left to right potatoe, peas, onion and 
sugarbeet. The classes can be separated with the 
parametric Bayes classifier for normal distributions. 
A test of the designed classifier on the same data 
as used for the design, produced a very high 
classification result, which is not surprising. 
However since we have no other data available, it is 
difficult to test the classification algorithm. 
Figure 7. Plot of May versus the projection feature 
(linear combination of the 2 July features). 
Figure 8. Histogram of the projection feature. Peaks 
are (left to right) potatoe, peas, onion and 
sugarbeet. 
To perform some sort of test, the data from fig. 2 
was automatically segmented using a split and merge 
algorithm (Ref. 5) and then classified. This brings 
a little variation in the data, because the field 
boundaries now differ from the ones in the manual 
segmentation. Of course, this is only a small effect, 
therefore care should be taken in the interpretation 
of the classifier results. Table 2 shows these 
results. The first 5 classes were used for optimizing 
the design of the classifier. Classes 6-8 represent 
a very small amount of data and cannot be considered 
to be representative. 
Class 8 (b 
these fiel 
easy to re 
not consid 
identified 
3 CONCLUSI 
In this pa 
of crop id 
improve th 
study by a 
enable ero 
growing se 
A hierar 
The succes 
separabili 
low grazin 
season (Ap 
other crop 
angular de 
35°. Field 
The test 
dataset as 
although fi 
different > 
should be 
results, s 
estimated 
Further 
ecological 
and more v; 
angular de] 
the Nether 
these subji 
4 ACKNOWLEl 
This study 
team. Part 
Sensing St< 
Contribuì 
Electronic! 
National A< 
collection, 
in Wageninj 
Theory Groi 
classifica) 
Mr. R. V] 
Electronics 
calculatioi