97 
computing the ratio between the true and classified forest 
border length in the independent reference sample area. 
RESULTS AND DISCUSSION 
For unsupervised classification, it was decided to limit the 
number of clusters to 119, because that is where a break 
occurs in the histogram of the initial clustered image. 
Table 5 shows the two decision trees that were 
successively applied to reclassify the result of 
unsupervised classification of the Landsat TM data. 
Figure 2 shows the final forest border map, which was 
obtained after the result of the second reclassification was 
thematically aggregated back into the 4 main classes and 
generalized with the sieve filter. 
A comparison of the land cover structure within the area 
of the reference aerial images (Table 3) shows the 
classified map to be closer to the true values than the 
CLC, especially for the "Forest" and "Shrub" classes. A 
site-specific comparison confirms the improvement in 
classification accuracy over the CLC database (Table 4). 
The thematic accuracy for the 4 main classes is estimated 
to 81,8% (Kappa 66,6%) for the rule-based classification 
and 75,3% (Kappa 57,3%) for the CLC database, while at 
the "Forest" / "Everything else" thematic level the 
accuracy increases to 91,2% (Kappa 81,3%) and 87,2% 
(Kappa 73,5%) respectively. It is evident from the error 
matrices (Table 4) that the accuracy is the lowest for the 
"Shrub" and "Abandoned pasture" classes. We attribute 
this to (1) their transitional character, making them difficult 
to consider in the decision trees and to (2) the lack of 
relevant information in the GIS layers. The forest border 
delineation in the classified map is slightly more accurate 
than the one in the CLC: the accuracy of the forest border 
delineation as estimated by the IREB value is ± 14 m for 
the classification and ± 15 m for the CLC. The minimum 
mapping unit is 0,25 ha for the classified map and 20 ha 
for the CLC. The classified map is therefore spatially more 
precise by definition. The improvement in precision is 
confirmed by the ratio of the classified to the true forest 
border length, which is 92,6% for the rule-based 
classification and 33,4% for the CLC database. 
It may come as a surprise that the rule-based 
classification performs so much better than the CLC, 
given that the rules/trees were learned by using the CLC 
as the target class. However, there is a logical explanation 
for this phenomenon, which has also been observed in 
other applications of machine learning and is known under 
the name of "clean-up effect" (Michie and Camacho 
1994). The learning process employed (which in our case 
also takes into account domain knowledge) has an 
averaging effect implicit in generalization which abstracts 
away the individual errors made by humans and yields 
performance similar to that of the trained humans but 
more dependable. 
Compared to the photointerpretation work on the 
Slovenian CLC database project (Kobler et al. 1998), 70% 
less man-days were needed to complete the classification 
of the study area. Proportionally less time should be 
needed for larger areas. 
True value 
Classification 
CLC database 
Forest 
62,1% 
62,8% 
57,6% 
Shrub 
5,0% 
6,6% 
11,9% 
Abandoned pasture 
9,8% 
7,8% 
9,9% 
Non-forest 
23,2% 
22,7% 
20,5% 
Table 3: Structure of the land cover within the area of the reference aerial images 
Classification 
CLC database 
Forest 
Shrub 
Aband. 
pasture 
Non 
forest 
Forest 
Shrub 
Aband. 
pasture 
Non 
forest 
TOTAL 
Reference 
data 
Forest 
29.072 
987 
536 
495 
26.779 
2.326 
504 
1.481 
31.090 
Shrub 
986 
862 
303 
334 
783 
1.027 
318 
357 
2.485 
Ab. pasture 
707 
963 
1.843 
1.394 
530 
1.625 
2.110 
642 
4.907 
Non-forest 
691 
497 
1.241 
9.169 
775 
1.002 
2.048 
7.773 
11.598 
TOTAL 
31.456 
3.309 
3.923 
11.392 
28.867 
5.980 
4.980 
10.253 
50.080 
Overall accuracy: 81,8% 
Kappa index of agreement: 66,6% 
Overall accuracy: 75,3% 
Kappa index of agreement: 57,3% 
Classification 
CLC database 
Forest 
Everything else 
Forest 
Everything else 
TOTAL 
Ref. 
data 
Forest 
29.072 
2.018 
26.779 
4.311 
31.090 
Everything 
else 
2.384 
16.606 
2.088 
16.902 
18.990 
TOTAL 
31.456 
18.624 
28.867 
21.213 
50.080 
Overall accuracy: 91,3% 
Kappa index of agreement: 81,3% 
Overall accuracy: 87,2% 
Kappa index of agreement: 73,5% 
Table 4: Comparison of the rule-based classification vs. the CLC database - error matrices and thematic 
accuracy assessment