All the input data were in a raster format, having 1880 x 
1360 pixels with a 25 m spatial resolution. The 
classification was divided into two successive phases 
(Figure 1): 
1. unsupervised classification of the multispectral 
satellite data and 
2. rule-based reclassification using also the GIS data. 
We tried to save as much effort as possible by automating 
certain procedures. This was done by using a machine 
learning algorithm, where the CLC database was the 
reference data set. The CLC database entirely covers the 
study area with a minimum mapping unit of 20 ha, giving a 
general overview of the land cover. It was assumed that, 
in spite of a considerable mapping generalization, some 
subjectivity problems and mistakes in the CLC database 
(Kobler and Hocevar 1999), there is enough information 
inherent in the CLC to use it as a reference both in the 
labeling stage of the unsupervised classification and in the 
machine learning stage of decision tree generation. The 
other, equally important, source of information when 
defining the decision tree was the domain expert 
knowledge. Before using the CLC as a reference, we 
aggregated the CLC nomenclature into the more general 
CLC_G classes relevant to our study. Because of the 
mixed land-use / land-cover nature of the CLC 
nomenclature, some of its classes were left out (denoted 
by 0 in Table 1). 
The Landsat TM satellite image (EC 1997) acquired in 
July 1995 had already been ortorectified previously (NLR 
1997) for the Slovenian CLC project. It was additionally 
radiometricaly corrected prior to classification, to alleviate 
the effects of variable illumination. Whereas the correction 
of atmospheric effect was skipped because of its 
complexity, the topographic normalization was performed 
using the Minnaert method (Smith et al. 1980), 
implemented in the SILVICS software (McCormick, JRC 
1999). A raster DEM (SMAS 1995a) with a 100 m 
resolution was used to derive the topographic variables. 
CLASSIFICATION METHODS 
The classification process encompassed unsupervised 
classification, followed by supervised classification, 
aggregation of the non-forest classes, and sieve filtering 
(Figure 1). The first phase - unsupervised classification - 
was used to group pixels in the Landsat TM radiometricaly 
corrected image channels 2, 3, 4, 5 and 7 into "natural", 
spectrally distinct classes. We decided against supervised 
classification at this stage because the spectral classes 
were so numerous, that it would be difficult to train on all 
of them. Each spectral class (i.e. cluster) was labeled 
according to the predominant CLC_G class. Visual 
examination of this first map approximation showed that 
there still remained some confusion among CLC_G 
classes in the output image. Some spectral classes 
related to more than one information (CLC_G) class, 
indicating that some information classes were spectrally 
similar and could not be distinguished from the 
multispectral data alone. 
During the second phase, additional information was used 
to derive two decision trees that successively improve the 
output of unsupervised classification. The additional 
information relates to the per-pixel values of different GIS 
layers, i.e. attributes (Table 2). The decision trees were 
generated by interactively combining domain expert 
knowledge with the results of automated induction of 
decision trees (i.e.machine learning). The machine 
learning of decision trees was based on the values of the 
CLC_G attribute. 
Decision trees (Quinlan 1986) predict the value (called 
class) of a discrete dependent variable from the values of 
a set of independent variables (called attributes), which 
may be either continuous (e.g. SLOPE) or discrete (e.g. 
FOREST81). Data describing a real system can be used 
to learn or automatically construct a decision tree. The 
common way to induce decision trees is the so-called 
Top-Down Induction of Decision Trees (TDIDT, Quinlan 
1986). Tree construction proceeds recursively starting 
with the entire set of training examples. At each step, the 
most informative attribute is selected as the root of the 
(sub)tree and the current training set is split into subsets 
according to the values of the selected attribute. For 
discrete attributes, a branch of the tree is typically created 
for each possible value of the attribute. For continuous 
attributes, a threshold is selected and two branches are 
created based on that threshold. For the subsets of 
training examples in each branch, the tree construction 
algorithm is called recursively. Tree construction stops 
when all examples in a node are of the same class (or if 
some other stopping criterion is satisfied). Such nodes are 
called leaves and are labeled with the corresponding 
values of the class. 
An important mechanism used to prevent trees from over 
fitting data is tree pruning. Pruning can be employed 
during tree construction (pre-pruning) or after the tree has 
been constructed (post-pruning). Typically, a minimum 
number of examples in branches can be prescribed for 
pre-pruning and confidence level in accuracy estimates 
for leaves for post-pruning. 
A number of systems exist for inducing classification trees 
from examples, e.g., CART (Breiman et al., 1984), 
ASSISTANT (Cestnik et al. 1987), and C4.5 (Quinlan, 
1993). Of these, C4.5 is one of the most well known and 
used decision tree systems. Its successor C5 (Quinlan 
1998) represents the state-of-the-art in decision tree 
induction at the time of writing this paper. The Windows 
implementation of C5, named See5, was used in our 
study. 
Out of 2.558.160 pixels in the study area a training subset 
of 127.537 pixels was selected for learning decision trees. 
To avoid bias, the training subset was selected in a 
stratified random fashion with equal number of pixels per 
stratum. The size of the subset was limited by the size of 
the smallest stratum. Two criteria were considered for 
stratification: the original CLC class, generalized into 8 
classes, and global yearly insolation (Gabrovec 1996), 
split into 2 classes at the median value. 16 strata were 
thus identified. To avoid possible problems at the 
landscape unit edges (mixed pixels and imprecise edge 
delineation in CLC), only pixels more than 100 m from the 
stratum edge were candidates for the training subset. 
95