93 
FOREST BORDER IDENTIFICATION BY RULE-BASED CLASSIFICATION OF LANDSAT TM AND GIS DATA 
Andrej Kobler and Dr. Milan Hocevar, Slovenian Forestry Institute, Slovenia 
Dr. Saso Dzeroski, Jozef Stefan Institute, Slovenia 
KEY WORDS: Rule-based classification, forest border, Landsat TM, GIS 
ABSTRACT 
The paper reports on a method for identification of the forest border and areas of spontaneous afforestation of 
abandoned farmland. Unsupervised and rule-based classification procedures are combined to classify Landsat TM and 
GIS data. The CORINE Land Cover database (CLC) database, one of the basic GIS layers in this study, lacks both 
spatial precision and thematic accuracy for environmental management applications at the regional to local scales. The 
study confirms the feasibility and efficiency of improving the precision and accuracy of the CLC forest border delineation 
by a sequence of automated and analyst-assisted classification procedures using the widely available data. 
Unsupervised classification is performed first, using CLC as reference in the labeling stage. A reclassification model is 
then developed combining domain expert knowledge with machine learning of decision trees, again using CLC as 
reference. Finally, the accuracy of the final map is estimated on an independent sample of photo-interpreted aerial 
photographs. Compared to CLC, the final reclassified map has slightly better accuracy and considerably higher spatial 
precision. This study is a part of a project aimed at developing an improved forest cover map of Slovenia as a foundation 
for a forestry information system. 
INTRODUCTION 
Slovenia has recently experienced an accelerated rate of 
infrastructural development and an increased rate of 
urban spread. Since the end of World War 2 Slovenia has 
also been undergoing a process of rapid deagrarization 
followed by a widespread spontaneous afforestation of 
abandoned farmland. Whereas there were only 36% of 
forests at the turn of the century (Zumer 1976), there are 
54 % of forests today according to official statistics (MAFF 
1997). The true percentage could be even as high as 
63,5% according to the most recent study of Kobler and 
Hocevar (1999). 
There is obviously an urgent need for suitable information 
for sustainable development to be used by decision 
makers in physical planning and environmental 
assessment. Demand seems to be the highest for 
accurate, precise and up-to-date land cover data, 
including GIS information regarding the forest area, forest 
type, forest landscape pattern etc. In the last decades 
there have been several successful efforts to map forest 
cover. Whereas some of them, like the land cover maps of 
the Kocevje forest management unit (Hladnik 1998), the 
Pomurska statistical region (Tretjak 1999) or the lower 
Vipava valley (Pavlin 1996) only tackled individual 
regions, others covered the entire country (EEA 1998, 
SMAS 1995, Zonta and SMAS 1999). The rapid process 
of afforestation, however, is continually making the 
existing forest maps obsolete. None of the conventional 
methods of large scale forest mapping used so far in 
Slovenia has been able to follow the rate of forest 
expansion. 
Although the recent satellite based CORINE Land Cover 
Slovenia (CLC) mapping project provided us with an up- 
to-date land cover map of the whole country (EEA 1998, 
Kobler et al. 1998), this map is still too generalized for 
users at regional to local scales (e.g. 1:25.000). This is 
due to the CLC methodology (EC 1993), which employs 
manual photointerpretation of 1:100.000 scale Landsat 
TM color prints. In addition, the CLC nomenclature 
contains some mixed classes, which make unambiguous 
forest border delineation from the CLC impossible. Some 
possibilities of automated updating of the CLC database 
have already been examined (Wilkinson et al. 1992) and 
there is an initiative for upgrading the CLC database to 
the 4 th thematic level (EEA PTL/LC 1998). However, at 
the time of writing this paper we found no studies 
published regarding automated improvement of this 
database in the sense of spatial precision. 
The main objective of this study is to develop a cost- and 
time-efficient method of automated forest mapping, using 
a combination of classification techniques. At the same 
time we aim to improve the classification accuracy and 
spatial precision of the CORINE Land Cover database 
forest border, which is one of the basic layers in our 
analysis. To ensure the applicability of the method beyond 
the study area, the method is based on widely available 
satellite and GIS data. Both unsupervised classification 
(clustering) and supervised classification (machine 
learning of decision trees) are used. The latter are further 
combined with domain expert knowledge to improve 
performance. The presented study is one stage in a 
broader project aimed at updating the forests map of 
Slovenia. 
STUDY AREA 
The study area covers 159.800 ha in the southwest of 
Slovenia, centered around the town of Postojna (Figure 
2). The area was chosen for its geographic and ecological 
diversity, thus raising issues of land cover classification 
representative for most of Slovenia. In the recent 
decades, the rural population has been mostly employed 
in the surrounding regional centers, abandoning farming 
and leaving farmland to spontaneous afforestation.