nbul 2004 
A HIERARCHICAL CLASSIFICATON OF LANDSAT TM IMAGERY 
FOR LANDCOVER MAPPING 
M.Avci®, Zuhal Akyurek® 
“General Command of Mapping, 06100 Cebeci Ankara, Turkey - mavci@hgk.mil.tr 
"Middle East Technical University, GGIT 06531 Ankara, Turkey - zakyurek(@metu.edu.tr 
  
WG IV/6 
KEY WORDS: Remote Sensing, Land Cover, Hierarchical Classification, Knowledge Base, Image, Multispectral, Spatial 
ABSTRACT: 
Information about current land-cover in forests is important for management and conservation of these areas. Up to the last decade 
traditional per pixel classification algorithms were used to be utilized in extracting land-cover information. However, they are poorly 
equipped to monitor land-cover in images acquired by current generation of satellite sensors with adequate accuracy. A good 
understanding and classification of an image can be done by gathering critical a priory knowledge about the study area and an 
effective use of channels involved in the procedure. It is important to make use additional spectral and spatial knowledge in order to 
improve the classification accuracy. In this study, a knowledge based hierarchical approach is proposed in order to classify and detect 
forest types in the Omerli Dam Lake Region. The method makes use of the fact that land-cover types and their associated knowledge 
form a natural hierarchy. Hierarchical classification is a powerful approach in solving classification problems by decomposing the 
image into a hierarchical tree structure. This also results in sub-dividing the area into spectrally consistent regions and helps dealing 
with spectral variability within each subarea. Three types of knowledge were involved in the rule-based classification of the study 
area: Domain spectral knowledge, Spectral classification rules obtained from training data and Spatial knowledge. Sub-dividing the 
area into smaller homogeneous regions in hierarchical classification increased the accuracy, while supervised classification technique 
yielded 47 per cent in the same area. Spatial reclassification involved in the hierarchical classification method increased overall 
accuracy, yielding new classes like coast. 
1.2 Background 
1. INTRODUCTION 
There are numerous accounts of research where TM data were 
1.1 Aim of the Study used to classify forest types, but few researchers have used a 
knowledge-based hierarchical approach. Anderson et al., (1976) 
Land-cover is one of the basic data layers in geographic Level I and II classifications (discrimination between deciduous 
information system for physical planning and environmental and coniferous forests) from remotely sensed data have been 
monitoring. Traditional multispectral image classification produced with accuracy of greater than 80 percent. However, 
techniques are, however, insufficient for extraction of land- for Anderson Level III classifications (discrimination between 
cover categories with required accuracy from high resolution forest types), mapping accuracy has been generally lower. Anil 
imagery. Attempts to increase the overall classification K. Jain (1989) proposed an image segmentation technique by 
accuracy, ranging from incorporation of ancillary data to use of extracting kernel information from the input image to provide 
expert systems and neural networks have proved to be hierarchical classifier to discriminate between major land-cover 
successful when compared with traditional classification types in the study area. A more detailed interpretation of the 
techniques. In this study, the classification accuracy problem image was then produced using a spatial clustering technique, 
was attacked using a knowledge-based hierarchical approach. In the previously extracted kernel image information and spectral 
the context of Landsat imagery, domain spectral knowledge, and spatial rules which make up the knowledge-base of the 
spectral classification rules obtained from training data and hierarchical classifier. 
spatial rules can be used to improve the quality of image 
classification (Anil, 1989). Since land-cover types present in A rule-based expert system was developed by Skidmore (1989) 
Landsat imagery form a hierarchical structure, a top-down to classify forest types. Relationships between forest type 
processing strategy was adopted in separate classes. classes and terrain (ie gradient, aspect and topographic position) 
were quantified using the knowledge of local forest personnel. 
The research described in this study attempts to deal with the The expert system had a higher mapping accuracy than the 
spectral variability within a landscape during image maps produced by traditional classifiers. Bolstad and Lillesand 
classification by subsetting large study areas into spectrally (1991) designed a system of programs named CLASSMOD 
consistent geographic areas with the cooperation of spectral and (Class Modeler) which allowed the integration of thematic data, 
spatial rules, classifying these areas independently and then satellite imagery and a rule-based, forward-chaining inference 
rejoining them to form a final continuous classification. In strategy in land-cover classification. Rules were used to 
essence, the hierarchical classification helps in reducing the describe feature types, data themes, the relationships among 
classification confusion among land-cover classes, because the themes and feature types and to define the inference path. 
spectral variability present within each subset image is usually Barnsley and Bar (1992) used a kernel-based reclassification 
considerably less than that between different strata. method, referred to as SPARK (Spatial Reclassification Kernel) 
which examined both the frequency and the spatial arrangement 
511