measurements in 1994 were to become familiar with the area, to 
collect some basic data for developing the interpretation method 
and to evaluate the possibilities to make field measurements on 
a large scale. The measurements in 1995 contain 259 points, 
half of which were used in the determination of rules and the 
other half in the estimation of the interpretation accuracy. 
2.3 Land use and class hierarchy 
The topographic map is composed of four major land use 
classes: water, forest, cultivated land and urban area. These 
classes include in practice the following areas: 
— Water: rivers, canals, fishponds and lakes, 
— Cultivated land: rice-, banana-, sugarcane- and vegetable 
fields, gardens, 
— Forest: different types of forest, 
— Urban: industry-. housing-, office-, traffic- and construction 
areas. 
The class hierarchy used in the interpretation is presented in 
Figure 1. Lowest in the hierarchy are 33 spectral classes which 
were interpreted in the Maximum Likelihood classification. 
Before the rule-based postclassification, these classes were 
combined so that 8 classes corresponding to the terminal nodes 
of the tree were obtained. The major land use classes are set 
nodes or terminal nodes in the tree. Some rules in the rule-based 
classification gave support to the terminal nodes and some 
others to the set nodes of the tree. 
  
Water bodies 
AN 
  
Land cover 
Vegetated land 
AN 
Non-vegetated land 
Ne 
  
  
Water Fishpond A Forest Urban area es 
Rice Garden Open Quarry 
wiw2 fpl..fpi r1... gl ga Tlgie wid wees Gyan 5g 
Figure 1. Land use and class hierarchy used in the interpretation. 
3. METHOD 
The interpretation method consists of segmentation, 
preclassification and rule-based postclassification of satellite 
and map data. In determining rules and believes for the rule- 
based classification. field measurements are used in addition to 
the data itself. Figure 2 shows the interpretation process. 
3.1 Segmentation 
Before interpretation, the satellite image is partitioned into 
spectrally homogenous, connected regions using a region-based 
segmentation method. Dealing with regions instead of pixels 
makes it possible to avoid noise in the interpretation result and, 
if needed, use spatial properties of the segments (for instance 
size, shape and neighbourhood) in addition to spectral 
information during the rule-based interpretation stage. 
A method based on hierarchical region merging (Beaulieu and 
Goldberg, 1989) was implemented and used to segment the 
image. Initial segments are obtained by merging neighbouring 
pixels which are similar enough using a threshold value for 
each band. After that a merging cost based on segment sizes 
and means of DN's is calculated for each neighbouring segment 
996 
pair in the image and the pair with minimum cost is merged. 
The merging costs are updated and merging is repeated until the 
desired number of regions is reached or the minimum merging 
cost exceeds a predetermined limit. For segmentation, the 
Zhong Shan image was divided into four subimages which were 
segmented separately and the results were then combined. 
Channels 3-5 of the TM image were used in the segmentation. 
The segmentation results have been evaluated visually and they 
are satisfactory. 
3.2 Preclassification 
The interpretation starts with the classical Maximum 
Likelihood method to interpret the segments based on training 
data. Channels 2-5 of the TM image were used in this study. 
The preclassification result contained 33 spectral classes, which 
were then combined to obtain the eight terminal classes in the 
class hierarchy, in order to start the next step in classification. 
The class probabilities obtained from the ML-classifier for each 
segment or the accuracy of the preclassification results can be 
used to determine belief for terminal node classes in the rule- 
based postclassification. In this study, the accuracy of the 
results compared to the reference points was used. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996