5 Japan —
a
vers, and sparsely
ng data due to the
1e advancement of
d uses from space.
veen image objects
al, and contextual
/HRS data directly
idy, we conducted
en image of a rural
ve extracted image
f image objects, a
d use classification
OF JAPAN
d use pattern often
on a panchromatic
A typical land use
;uses (usually with
etc.), forests, and
vn in Japan tend to
id form residential
ded by agricultural
c. The boundaries
iclds are relatively
; are not in a grid-
of parcels are not
ackyard or a small
t is, image objects
ral fields in rural
tuents.
:ntial areas can be
use pattern in the
-pixel-base image
use maps without
S.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Agricultural Field
Figure 1. Land use pattern of the study area
(a) A land use pattern of a rural town in Japan depicted on
an IKONOS (* 3SI) image. (b) Close-up of a residential
area, (c) Close-up of an agricultural field.
3. OBJECT-ORIENTED APPROACH
Providing the complexity of the land use pattern in rural parts in
Japan, traditional pixel-by-pixel-base image classifiers may
hardly produce optimal land use maps from VHRS images
without intensive editing by human photo-interpreters, due to
the similarity in the land cover materials at the order of meters
among different land use classes in those areas. To cope with
the problem, contextual information such as geometrical
properties of image objects including the sizes, shapes, etc. and
spatial relationships among them should be incorporated in the
mapping process as a set of classification rules. Object-based
classifiers combining image segmentation and contextual-rule-
based labelling are expected to be a useful tool for land use
mapping from VHRS data.
In the past two decades, various image segmentation techniques
have been developed to incorporate context in the image
classification procedure (van der Sande et al. 2003). In this
study, a segmentation technique developed by Baatz and
Schüpe (2000), which is a type of region growing multi-scale
segmentation algorithms, was used to study object-oriented land
use classification.
Major research topics of this study are twofold:
- To study correspondence between image objects at
different scales and geographic features. How to optimize
segmentation results for mapping rural land uses.
- To select contextual information that can be used as a set
of rules for contextual labelling of land use classes on the
image objects produced with the image segmentation.
4. IMAGE OBJECTS AT DIFFERENT SCALES AND
CORRESPONDING GEOGRAPHIC FEATURES
In this study, a set of image segmentation using the algorithm
developed by Baatz and Schäpe (2000) with different scale
parameters were conducted using an IKONOS pansharpen data
(4 band) of a rural town in Kouchi-ken in south-western Japan
taken on November 21, 2001 to study the correspondence
between image objects at different scales and geographic
features. The segmentation criterion used in this study for
merging multiple image objects into a larger image object was a
standard deviation of the pixel values in an image object
(segment).
Figure 2 shows two examples of the subset of the segmentation
results with an original IKONOS data; these are at the
intermediate steps corresponding to certain scale parameters of
a region growing process started from individual pixels.
Although they do not perfectly match the shapes of geographic
features on the ground, image objects produced with certain
scale parameters shows relatively good correspondence with
some geographic features. Figure 2 (b) shows the segmentation
result with scale parameter 75, and the image objects depicted
with white lines correspond to the rooftops of individual houses
and parcels of agricultural fields relatively well. On the other
hand, image objects in Figure 2 (c) appeared at scale parameter
350 correspond to larger geographic features such as the
boundary between residential areas and agricultural fields.
Figure 3 and Figure 4 show examples of the growth curves of
the areas of image objects to exemplify how an image object,
starting from an individual pixel corresponding to a rooftop of a
house, a tree crown, a agricultural field, etc. increases its area in
accordance with the increase of the scale parameter. Studying
the optimization of image segmentation, Usuda et al. (2003)
focused on the stable periods observable in growth curves of the
areas of image objects as a key to decide optimal scale
parameters for specific applications of image segmentation
techniques. Stable periods are thought to be 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. Stable periods of the
growth of image object sizes are also observed in the image
segmentation experiments in this study. In this study, every
growth curve of an image object corresponding to a certain
gcographic feature has multiple stable periods. In addition, the
change from one stable period to the next stable period on a
growth curve is usually abrupt. These results suggest there is a
hierarchical spatial structure in the land use of the study area.
Based on the analysis of the growth curves of the image objects
produced with the multi-scale image segmentation, each land
use class appears to have a class-specific growth pattern (see
Figure 3 and Figure 4). Considering the correspondence
between the stable periods appearing in the growth curves and
