International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
v ITE : 
shows the segmentation result at scale parameter 20. Many 
image objects at this scale correspond to individual tree crowns, 
as well as other spatial objects including roofs of greenhouses, 
and the ground sizes of most image objects at scale parameter 
20 are smaller than the typical patch sizes of the vegetation 
(land cover) classes shown in Table 1. On the contrary, image 
objects produced at scale parameter 300 appear to correspond to 
much coarser units of spatial features. Image objects shown in 
Figure 1 (c) at scale parameter 80 match polygons of the land 
cover classes shown in Table 1 relatively well. A research 
question is how to determine an optimal set of segmentation 
parameters for vegetation mapping in rural areas in Japan. 
Usuda et al. (2003) focused on the stable periods observable in 
a growth curve of the sizes of image objects during a region 
growing process as a key to decide optimal scale parameters for 
specific applications of image segmentation techniques. Stable 
periods are the period when standard deviations of pixel values 
of neighbouring image objects are apparently different, and the 
merge process of the image objects becomes relatively slow. 
Such stable periods of the growth of image object sizes are also 
observed in the image segmentation experiments in this study. 
Each class of vegetation classes appears to have an optimal 
scale parameter to be delineated properly. 
5. SEGMENT-BASED CLASSIFICATION 
In the object-based image classification, each segmented image 
object needs to be labelled with a proper class name. Some 
image objects in this study were relatively easy to be labelled 
based on the spectral properties, and others are difficult to be 
determined only by the spectral properties without considering 
other contextual information such as relative sizes, spatial 
relationships, texture, and so on. The IKONOS data used in this 
study was taken in November, and vegetation-covered and non- 
vegetation areas were spectrally distinctive on the imagery. 
Among the vegetation-classes, grasses and deciduous forests 
showed relatively distinctive spectral properties on the 
IKONOS imagery so that their labels were reliably assigned 
based on their spectral information such as averages and 
variances of pixel values belonging to specific image objects. 
On the contrary, instances of some vegetation classes were 
difficult to distinguish from other vegetation classes. For 
instance, bamboo forests and other trees (both coniferous and 
deciduous) showed very similar reflectance properties in the 
visible and near-infrared spectral bands so that other properties 
of the image objects must be used for the labelling rules. For 
distinguishing bamboo forests from coniferous trees, texture 
information of image objects appears to be useful. Figure 2 
shows the textural difference between bamboo and coniferous 
tree areas. Variances of pixel values in 3 x 3 moving windows 
are tend to be larger at pixels corresponding to bamboo forests 
than those of other trees. 
Sizes and shapes of image objects were also useful properties 
for distinguishing some classes such as rice fields from others. 
288 
  
¥ t 
Figure 2. Texture analysis of deciduous tree- and 
bamboo-covered areas 
Above: IKONOS (* 3SI) imagery 
Below: Texture image 
(Variance in 3 x 3 pixels) 
A: Bamboo (brighter area) 
B: Deciduous tree (darker area) 
6. PIXEL-BASED CLASSIFICATION 
Traditional pixel-based image classifiers such as the maximum 
likelihood classifier tend to produce salt-and-pepper-like 
classification results representing land covers of the study arca 
in the aperture size of the ground resolution of the remote 
sensing data. These types of salt and peppers are usually seen as 
noises for land cover mapping in satellite remote sensing 
analyses. In the case of vegetation mapping using very high- 
resolution satellite data, salt-and-pepper-like classification 
results are useful in combination with object-oriented 
classification results, because pixel-based classification results 
provide additional information about the distribution and 
density of vegetation covers in the image objects delineated by 
object-based classifiers. 
Figure 3 compares three vegetation maps obtained using 
maximum likelihood classifier, object-oriented classifier, and 
aerial photo-interpretation. 
  
Internat 
  
This paps 
data and 
and pixc 
developm 
provide a 
satellite r 
informati 
analysis 1 
study sug. 
informatic 
environm 
Reference 
Baatz, M. 
- an opt 
segmen 
Informa 
T and 
12-23. 
Ministry © 
Bureau, 
Census 
Usuda, Y. 
segment 
of the