International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
  
(a) IKONOS imagery 
(b) Segmentation Result 
(* 3SI) (Scale parameter 75) 
(c) Segmentation Resul 
(Scale parameter 350) 
Figure. 2 Image Segmentation results 
9090909009000900009 
   
   
   
0600000000000 
......d.0.. 
99900600 
© 9 99 9 $00 090900900909 © © 
.... 
  
oo 20900009909 0090990 09 0 0 00 9 0 00 O © © 0 0 0990000 090000 009009 € »e 
eo 0 00 000.000 
Figure. 3 A growth curve of the area of an image 
object starting from a pixel belonging to a house 
corresponding image objects (segments), the segmentation 
results at the stable periods seem to reflect the hierarchical 
spatial structure of the study area. For instance, a small-scale 
pattern in the hierarchical spatial structure corresponds to the 
rooftops of houses. Parcels of houses appear to match the image 
objects at the next level of the spatial structure. Further, the 
polygons corresponding to the boundary between residential 
areas and agricultural fields seem to match the larger spatial 
pattern. However, the larger the image objects become, the 
weaker the correspondence between the stable periods and land 
use classes becomes. Moreover, the length of stable periods 
does not appear to be an appropriate index for obtaining optimal 
segmentation results, based on the observation of the 
correspondence between the lengths of the stable periods and 
segmentation results; the length of the stable periods does not 
correspond to meaningful land use classes such as the apparent 
boundary between residential areas and surrounding agricultural 
fields. Instead, stable periods appear in the early stages of the 
growth curves, i.e. small-scale image objects, show relatively 
good geometrical correspondence with basic land cover patches 
such as rooftops of houses, patches of grasses, plots of farmland, 
etc. 
oo... 
    
    
........“ 
9999909 
eo. 
  
ee 00000. 0 0.0000.00. 0. 0900000000000 005099 00000 050000 À 
e eo 000. 000 0006000000000000v 
Figure 4 A growth curve of the area of an image object 
starting from a pixel belonging to an agricultural plot 
5. LAND USE CLASSIFICATION FRAMEWORK 
Figure 5 shows a schematic of the land use classification 
process applying contextual rule-based labelling techniques to 
the segmentation results. 
According to the analysis in the previous chapter, segmentation 
results at a stable period in the early stage of the region growing 
match basic land cover patches relatively well. These basic land 
cover units include small patches of bare land, grass, parts of 
manmade structures such as rooftops, trees, open water, etc. 
These basic land cover patches are common constituents of 
most land use instances. For example, small patches of bare 
land may be a part of a backyard of a house, or a part of a 
fallow. In the proposed classification framework, labelling of 
land use classes start from small image objects produced at an 
early stable period which correspond to the basic land cover 
classes at small scale. 
Sizes and spatial relationships of the image objects would be 
the next criteria to determine labels of image objects. Relatively 
isolated small patches of bare soil may be part of backyards of 
   
Internat 
houses. 
may be 
Spectra 
detailed 
classific 
deciduo 
texture 
À land 
the abo 
an exist 
of aeria 
use ma] 
good co 
This stu 
produce 
use cla: 
through 
Suggest: 
object-c 
propose 
use ma 
and gen